A16z internal data this week put 29% of the Fortune 500 as live, paying customers of a leading AI startup. OpenAI reported enterprise customers now exceed 40% of revenue, on pace to match consumer by end of 2026. Fortune 500 companies historically took years to land a first enterprise contract; the usual ramp didn’t happen this time.

Two companies documented what company-wide deployment looks like from the inside. Ramp, at 99% AI tool adoption, found most employees still couldn’t configure their own setups, so they built Glass, an internal system where one person’s workflow breakthrough becomes a shared organizational capability automatically. Every gave each employee a personal AI agent and within two months it had changed everything about how they work, surfacing challenges the field hasn’t named yet, including the ant death spiral, where agents loop in conversation only with each other.

Share

Boston University economist James Bessen reports developer headcount is up 400,000 jobs since ChatGPT launched, with AI expanding the software market faster than it displaces workers, a pattern that held across every prior wave of productivity gains. This update also fills a gap from late March: Stanford’s James Zou built a virtual lab of AI scientist agents, and found that agents working as a team generate more creative hypotheses than single agents working alone.

Apr 9, 2026 · Ramp
After hitting 99% AI tool adoption, Ramp found most employees couldn’t configure their own setups. They built Glass, an internal AI productivity suite, to make every worker an AI power-user without technical friction. The system propagates one person’s workflow breakthroughs as shared organizational capabilities, with the product itself serving as the enablement rather than workshops or training.

Apr 8, 2026 · Every
Every gave each employee a personal AI agent (OpenClaw), and within two months it had completely changed everything about the way they work. Agents adopt their owners’ professional identity inside Slack, building trust as proxies for their humans. The team surfaced real deployment challenges: memory gaps, group chat etiquette, and what they call the ant death spiral, where agents loop in conversation only with each other.

Apr 8, 2026 · (Multiple)
A16z internal data shows 29% of the Fortune 500 and 19% of the Global 2000 are already live, paying customers of a leading AI startup. That penetration rate is unprecedented for early-stage tech. Fortune 500 companies historically took years to land a first enterprise contract; AI skipped that ramp. The finding pushes back against a widely-cited MIT study claiming 95% of generative AI pilots fail.

Apr 7, 2026 · Anthropic
Anthropic’s unreleased Claude Mythos Preview has found thousands of high-severity vulnerabilities across every major operating system and browser, surpassing “all but the most skilled humans” at finding and exploiting software flaws. Project Glasswing enlists AWS, Apple, Google, Microsoft, and a dozen other tech leaders to deploy the model defensively, with Anthropic committing $100M in compute credits to secure critical open-source infrastructure.

Apr 7, 2026 · James Bessen
A TPRI report from Boston University economist James Bessen finds that software developer headcount has added 400,000 jobs since ChatGPT launched, with output per developer rising roughly 6% annually. AI is expanding the software market faster than it displaces workers, enabling demand for new products. The same pattern held across every prior wave of developer productivity gains: structured coding, DevOps, cloud.

Apr 2, 2026 · OpenAI
OpenAI’s Chief Revenue Officer reports enterprise customers now exceed 40% of revenue, on track to match consumer by end of 2026. Codex reached 3 million weekly active users and APIs handle 15 billion tokens per minute. New customers include Goldman Sachs, State Farm, and Philips. OpenAI is positioning AI as a company-wide intelligence layer, replacing individual copilots with coordinated agents across entire businesses.

Mar 31, 2026 · James Zou (Stanford)
Stanford professor James Zou built a virtual lab of AI scientist agents that mirrors his physical research team, with an AI professor and student agents assigned specialties in immunology, cardiology, and data science. The agents hold group meetings, run experiments on a budget, and self-train by reading papers and passing quizzes. Zou found AI teams generate more creative hypotheses than single agents working alone.

Subscribe now

Read the full log (over 100 entries now!) at

Updated Sun Apr 12, 2026, 11:00 AM EDT

Welcome to the Treeline! The view clears here. Enjoy a quiet place to read without distraction.

This article captures AI’s transformation of work, the good and the bad. The full log of all entries follows the summary. Subscribe for weekly updates.

Share

On March 10, 2026, agents created more cells in Hex than humans did for the first time, a threshold that CEO Barry McCardel said arrived faster than expected. The same week, Amazon called an emergency meeting over four severe site outages attributed to “Gen-AI assisted changes.” Before the meeting, its engineers had quietly scrubbed the internal memos tracking the trend since Q3 2025. AI-assisted code had already caused hours of downtime, the failures compounding across millions of transactions.

OpenAI reported in April 2026 that enterprise customers exceed 40% of revenue. Codex reached 3 million weekly active users, with APIs handling 15 billion tokens per minute. A16z data shows 29% of the Fortune 500 and 19% of the Global 2000 are paying customers of a leading AI startup, a penetration rate unprecedented for early-stage technology. Anthropic closed a $30 billion Series G in February 2026, valued at $380 billion on $9 billion in run-rate revenue. Claude Code had already crossed $1 billion by December 2025, while most cloud categories were still consolidating around a dominant vendor.

Software engineering is the clearest leading indicator of AI’s impact on knowledge work, the field where transformation compounds fastest. Cursor’s co-founder reported in February 2026 that cloud agents now create more than a third of the company’s merged pull requests, running autonomously in isolated VMs on long-horizon tasks. He predicted the majority of development work will shift to agents within a year. Stripe’s unattended agents produce over 1,000 human-reviewed pull requests each week, launched from a single Slack message. No hand-written code is involved. An OpenAI team shipped a complete product the same way, one million lines across 1,500 pull requests in five months.

Andrej Karpathy named December 2025 the moment coding agents crossed a hard threshold. Simon Willison, Django’s co-creator, placed the inflection point a month earlier. AI has broken his ability to estimate project timelines, he said, after 25 years of building software. Customer outcomes confirmed the shift. Wiz migrated 50,000 lines of Python to Go in roughly 20 hours, a project estimated at two to three months of specialized engineering. Rakuten cut time to market 79 percent using Claude Code across its engineering organization. Addy Osmani, documenting the structural shift at Google in March 2026, put it plainly: the agent, not the file, is now the unit of work.

The displacement side is accelerating. An NBER working paper drawing on 750 CFOs projects 502,000 AI-driven job cuts in 2026, nine times the 55,000 cut in 2025. White-collar and entry-level roles are disappearing fastest. In some firms, agents already assign work to humans and inform decisions about promotions and layoffs. Oracle shed roughly 10,000 employees in late March 2026, cuts described as not performance-based. Block cut nearly half its workforce in February; Atlassian followed in March, eliminating 1,600 roles across ten percent of its global staff.

Amazon has cut more than 30,000 workers since October 2025. ADP Research, surveying 39,000 workers across 36 countries, found only 22% strongly agreed their job was safe. Multiple analysts dismiss these as standard cost-cutting dressed in AI language. Independent sources point to structural pressure on entry-level white-collar work, consistent across enough methods to indicate lasting change.

The productivity evidence has seemed rigorous, though recent findings complicate the picture. The Noy and Zhang experiment, published in Science in July 2023, found AI raised professional writing productivity substantially, with the largest gains flowing to weaker workers and narrowing skill gaps within occupations. A comprehensive ICLE review covering writing, software, legal, accounting, and customer support found consistent gains of 15 to 50 percent. A QJE study of 5,172 customer support agents found a 15% average increase, with international workers improving their English fluency and customers becoming measurably more polite. Goldman Sachs documented a 20% productivity gain for coders and a 15% drop in post-release bugs across more than 46,000 employees. The BCG and Harvard study of 758 knowledge workers coined “jagged technology frontier.” AI users completed 12.2% more tasks, 25.1% faster, for work inside the frontier. For complex managerial tasks outside it, users were 19% less likely to produce correct solutions. MIT researchers found that AI-generated code creates technical debt surfacing months later, potentially erasing the gains. Amazon’s four site outages in a single week showed what happens when deployment outpaces best practices. ADP found some workers report task time increasing up to 346% when adoption proceeds without workflow redesign.

Below the headline numbers, adoption has reached every sector touching knowledge work, some visible in press releases, some only in restructured hiring. Klarna’s AI assistant handled 2.3 million conversations in its first month in February 2024, doing the work of 700 full-time agents. By January 2026, McKinsey had deployed 20,000 AI agents and added AI aptitude to its graduate hiring criteria. Walmart built four “super agents” to unify experiences across customers, associates, partners, and developers. The FDA had cleared 873 AI algorithms for medical imaging by mid-2025, 115 in that year alone, with radiology as the single largest clinical target. Accounting firms are restructuring junior pipelines toward “digital seniors” who oversee AI workflows. A two-person law firm now handles workloads of much larger practices. Economist Scott Cunningham generated a complete academic paper in under two hours for roughly $100, covering idea, data, analysis, referee responses, and code audit. Stanford HAI documented inference costs falling 280-fold since 2022. METR’s benchmarks show task-completion horizons growing exponentially. OpenClaw went from zero to 250,000 GitHub stars in under four months, overtaking React and Linux on the all-time leaderboard.

IBM is tripling entry-level hiring across engineering, HR, and support. Companies that hollow out junior pipelines now, IBM argues, will struggle in five years. That announcement landed in February 2026, the same month Block cut nearly half its staff.

Log

Apr 9, 2026 · Ramp
After hitting 99% AI tool adoption, Ramp found most employees couldn’t configure their own setups. They built Glass, an internal AI productivity suite, to make every worker an AI power-user without technical friction. The system propagates one person’s workflow breakthroughs as shared organizational capabilities, with the product itself serving as the enablement rather than workshops or training.

Apr 8, 2026 · Every
Every gave each employee a personal AI agent (OpenClaw), and within two months it had completely changed everything about the way they work. Agents adopt their owners’ professional identity inside Slack, building trust as proxies for their humans. The team surfaced real deployment challenges: memory gaps, group chat etiquette, and what they call the ant death spiral, where agents loop in conversation only with each other.

Apr 8, 2026 · (Multiple)
A16z internal data shows 29% of the Fortune 500 and 19% of the Global 2000 are already live, paying customers of a leading AI startup. That penetration rate is unprecedented for early-stage tech. Fortune 500 companies historically took years to land a first enterprise contract; AI skipped that ramp. The finding pushes back against a widely-cited MIT study claiming 95% of generative AI pilots fail.

Apr 7, 2026 · Anthropic
Anthropic’s unreleased Claude Mythos Preview has found thousands of high-severity vulnerabilities across every major operating system and browser, surpassing “all but the most skilled humans” at finding and exploiting software flaws. Project Glasswing enlists AWS, Apple, Google, Microsoft, and a dozen other tech leaders to deploy the model defensively, with Anthropic committing $100M in compute credits to secure critical open-source infrastructure.

Apr 7, 2026 · James Bessen
A TPRI report from Boston University economist James Bessen finds that software developer headcount has added 400,000 jobs since ChatGPT launched, with output per developer rising roughly 6% annually. AI is expanding the software market faster than it displaces workers, enabling demand for new products. The same pattern held across every prior wave of developer productivity gains: structured coding, DevOps, cloud.

Apr 3, 2026 · Asanify
Asanify placed an autonomous AI named Ivy into a live Junior HR Executive role with Slack access, a company email, HR admin credentials, and KPIs matching those of a human hire. Ivy handled real employees with real consequences over seven days. The company framed it as “eating its own cooking.” They build AI-native HR software and needed to know whether their technology could genuinely perform an entry-level role, not merely assist one.

Apr 2, 2026 · Simon Willison
Software engineer and Django co-creator Simon Willison described AI’s impact on engineering as a structural shift in an April 2026 podcast appearance. He called November 2025 an inflection point, named software engineers as bellwethers for all knowledge workers, and said AI has broken his ability to estimate project timelines. He flagged dark factories, fully AI-run operations, as an emerging model, and observed the transition is especially hard for mid-career engineers.

Apr 2, 2026 · OpenAI
OpenAI’s Chief Revenue Officer reports enterprise customers now exceed 40% of revenue, on track to match consumer by end of 2026. Codex reached 3 million weekly active users and APIs handle 15 billion tokens per minute. New customers include Goldman Sachs, State Farm, and Philips. OpenAI is positioning AI as a company-wide intelligence layer, replacing individual copilots with coordinated agents across entire businesses.

Mar 31, 2026 · Oracle
Oracle let go roughly 10,000 employees in late March 2026, eliminating roles for senior engineers, architects, operations leaders, program managers, and technical specialists. The company declined to comment on the cause, but executives had previously stated that AI tools “enabled fewer employees to do more work.” The cuts were described as not performance-based, signaling a structural workforce reduction tied to the company’s heavy AI investment.

Mar 31, 2026 · James Zou (Stanford)
Stanford professor James Zou built a virtual lab of AI scientist agents that mirrors his physical research team, with an AI professor and student agents assigned specialties in immunology, cardiology, and data science. The agents hold group meetings, run experiments on a budget, and self-train by reading papers and passing quizzes. Zou found AI teams generate more creative hypotheses than single agents working alone.

Mar 30, 2026 · INSEAD/HBS
A randomized controlled trial across 515 global startups found that helping firms discover where to deploy AI produced dramatic firm-level gains: treated startups completed 12% more tasks, were 18 percentage points more likely to acquire paying customers, and generated 1.9x higher revenue, while reducing their need for external capital by 39.5%. Each additional AI use case discovered added ~26% more revenue.

Mar 27, 2026 · Zach Lloyd (Warp)
Warp CEO Zach Lloyd details how his company stopped buying SaaS, replacing subscriptions with agents, skills, and just-in-time apps. They migrated 266 documentation pages from a hosted platform to Markdown in hours, replaced community-insight SaaS with a monitoring agent that suggests replies and self-improves, made data analytics self-serve via agent skills with BigQuery access, and built custom recruiting tools from raw APIs. Savings exceeded $10k per year on docs alone.

Mar 27, 2026 · Tesla
Tesla this year phased out car models, including a popular luxury sedan, to stand up a new Optimus humanoid production line. Elon Musk, who frames the project as building a world of abundance where work is optional, merged SpaceX’s AI unit with Tesla’s robot program. At least three other major tech investors are funding competing humanoid programs. The push represents a direct bet that physical AI, not just software, will make human labor obsolete.

Mar 26, 2026 · (Multiple)
An NBER working paper drawing on 750 CFOs found AI-driven job cuts in 2026 are projected at 502,000 — nine times the 55,000 cut in 2025. ADP Research, surveying 39,000 workers across 36 countries, found only 22% strongly agreed their job was safe. A counterintuitive finding: workers using AI tools report time on some tasks has increased up to 346%, as adoption without workflow redesign creates friction rather than speed.

Mar 25, 2026 · Basecamp
Basecamp announced full CLI and agent access to its platform, letting AI agents do anything a human can — write documents, manage to-dos, answer check-ins. Claude, Codex, OpenCode, and Cursor can all connect directly. The move reframes project management software as agent-native infrastructure, positioning the user’s AI tooling at the center rather than the app’s own interface.

Mar 24, 2026 · Anthropic
Anthropic’s latest Economic Index finds that 49% of jobs now have at least a quarter of their tasks performed using Claude, up from a narrower set of high-value coding tasks in late 2025. Usage is diversifying toward lower-wage occupations as adoption broadens. The report’s central finding: experienced users attempt higher-value tasks and elicit more successful responses, suggesting early AI adoption creates self-reinforcing skill advantages that could widen inequality between early and late adopters.

Mar 21, 2026 · Addy Osmani
Google engineer Addy Osmani documents a structural shift in developer tools: the primary work surface is moving from the code editor to an agent orchestration layer. Cursor’s new Glass interface, GitHub Copilot Agents, Conductor, and Claude Code Web all converge on the same pattern — developers specify intent, delegate to parallel agents running in isolated workspaces, then review diffs. The agent, not the file, is becoming the unit of work.

Mar 21, 2026 · Azeem Azhar (Exponential View)
Azeem Azhar analyzes Jensen Huang’s GTC declaration that every company needs an OpenClaw strategy. The core argument: the AI debate fixates on capability when diffusion is what matters. The harness — not the engine — is the revolution, just as the automobile mattered more than the combustion engine alone. Tokens are now a productive input as fundamental to knowledge work as electricity. Organizations treating token budgets as an IT cost center rather than a business input are dangerously behind.

Mar 20, 2026 · (Multiple)
Q1 2026 has already seen more tech layoffs than Q1 2025, with Amazon, Meta, Salesforce, and Block among companies citing AI. But analysts are calling it: much is standard cost-cutting dressed in AI language. Block’s Jack Dorsey cut nearly half the workforce while ‘embracing AI,’ prompting critics to call the framing ‘convenient.’ The Federal Reserve Bank of St. Louis found a striking correlation between AI prevalence and unemployment increases since 2022, but a Harvard Business Review report suggests many layoffs are driven by AI’s potential, not its performance.

Mar 20, 2026 · Armin Ronacher
Armin Ronacher, creator of Flask and a developer at the center of AI tooling, reports that time saved by AI gets immediately captured by competition: anyone who pauses is outmaneuvered by someone filling every freed-up hour with new output. A YC batch last year saw startups vanish without notifying customers, treating proper shutdowns as wasted time. Ronacher argues the pressure to ship at inference speed is eroding the trust, community, and long-term commitment that durable software depends on.

Mar 19, 2026 · Hex
On March 10, 2026, agents created more cells in Hex than humans did for the first time. Hex CEO Barry McCardel announced the milestone on LinkedIn: in a product where queries, charts, and code are the atomic units of work, the balance tipped from human-created to agent-created. Usage has gone parabolic, McCardel wrote, and the transition happened faster than expected as agentic surfaces became deeply embedded in core workflows.

Mar 19, 2026 · Astral
OpenAI is acquiring Astral, the company behind Python tools Ruff, uv, and ty, collectively reaching hundreds of millions of downloads monthly. Astral will join OpenAI’s Codex team. CEO Charlie Marsh framed the move as working at ‘the highest-leverage thing we can do’ at the frontier of AI and software. The acquisition consolidates foundational Python development tooling inside the AI company now building coding agents intended to automate engineering work.

Mar 17, 2026 · Intercom
Intercom engineer Brian Scanlan detailed how the company built a full-stack engineering platform on Claude Code: 13 plugins, 100-plus skills, and automated hooks. The most striking component is a read-only Rails production console via MCP, giving Claude live access to execute arbitrary Ruby against production data for feature flag checks, cache inspection, and business logic validation, with Okta authentication and blocked critical tables as safety gates.

Mar 12, 2026 · Oracle
Oracle raised its restructuring fund to $2.1 billion and is cutting thousands of jobs, but Fast Company’s analysis found the bigger driver is the company’s $50 billion AI data center buildout rather than AI directly replacing workers — a distinction that applies to many tech layoffs being framed as AI-driven.

Mar 12, 2026 · Notion
Notion co-founder Simon Last describes how engineering teams are reorganizing around AI agents, with developers moving from individual code authors to managers of multi-agent systems that generate, review, and fix each other’s output with limited human intervention.

Mar 12, 2026 · (Multiple)
An extended NYT Magazine portrait of software engineers who barely write code anymore. Developers now spend their days directing AI agents in plain English. At Hyperspell, a task that once took a full day takes 30 minutes. Prompt files have become behavioral contracts — including stern commands like ‘pushing code that fails pytest is unacceptable and embarrassing’ to shape AI output. The piece documents a shift already underway: engineer as manager.

Mar 11, 2026 · Atlassian
Atlassian cut 1,600 roles, 10 percent of its global workforce, in response to AI’s effect on how software is built. Co-founder and CEO Mike Cannon-Brookes framed it as choosing to adapt rather than a cost-cutting emergency. Australia’s largest listed technology company joins a list of enterprise software firms shedding staff as AI tools compress engineering output, shrinking the headcount needed to sustain development at scale.

Mar 11, 2026 · Boston Consulting Group
In a preregistered experiment with 758 BCG knowledge workers, AI assistance completed 12.2% more tasks, 25.1% faster, with higher quality — on tasks within AI’s capability frontier. But for a complex managerial task outside that frontier, AI users were 19% less likely to produce correct solutions. The HBS/BCG study coined ‘jagged technology frontier’ to describe how AI’s impact varies unpredictably even across seemingly similar tasks in the same knowledge workflow.

Mar 10, 2026 · Amazon
Amazon called an emergency engineering meeting after four severe site outages in a week. An internal document cited “Gen-AI assisted changes” as a “trend of incidents” since Q3, though the reference was quietly deleted before the meeting began. Amazon said the incidents had a “high blast radius.”

Mar 10, 2026 · Amazon
Amazon’s ecommerce leadership called an emergency all-hands after a ‘trend of incidents’ linked to ‘Gen-AI assisted changes’ — including a six-hour site outage that left customers unable to complete transactions. A company memo cited ‘novel GenAI usage for which best practices and safeguards are not yet fully established.’ Junior and mid-level engineers now require senior sign-off on AI-assisted changes. Amazon Web Services suffered separate AI-related incidents in the same period.

Mar 8, 2026 · Aaron Levie (Box)
Box CEO Aaron Levie argues that agents will become the primary users of software, with enterprises eventually fielding 100 times more agents than people. Every contract review, financial audit, customer support case, and line of code will flow through agents, making software designed for humans effectively obsolete.

Mar 8, 2026 · (Multiple)
As companies like Snowflake deploy AI agents that handle monitoring, on-call response, and task assignment, hierarchies are reshaping: agents now assign work to humans and, in some firms, inform decisions about promotions and layoffs. White-collar jobs, especially entry-level roles, are disappearing fastest.

Mar 7, 2026 · Benzinga
Salesforce’s CEO dismissed AI layoff fears on Wednesday; by Friday the economy had shed 92,000 jobs and Morgan Stanley, Oracle, and Capital One had all announced cuts, complicating his argument.

Mar 6, 2026 · (Multiple)
Even from Management* — A Udacity survey found only 9% of executives and managers want to replace their entire workforce with AI. Most cite human creativity, customer preference for human interaction, and legal liability as limits AI can’t clear. Companies rushing to cut headcount in favor of agents are likely to meet resistance across the org chart.

Mar 5, 2026 · Oracle
Oracle is planning thousands of layoffs across multiple divisions, some targeting categories expected to shrink due to AI, as the company faces a cash crunch from its massive AI data center expansion.

Mar 5, 2026 · Alex Imas
A living review of micro and macro evidence finds AI productivity gains, long documented in individual studies, are now appearing in aggregate economic data for the first time.

Mar 5, 2026 · Cursor
Forbes profiles Cursor, the fastest-growing AI coding company, as it confronts an uncomfortable question: if AI agents can write code autonomously without a human code editor, what is a code editor for? Cursor grew explosively as a developer tool but now faces pressure from pure agent products that bypass editors entirely. The company is building deeper agentic capabilities, competing in a landscape its own success helped create.

Mar 5, 2026 · Texas Society of CPAs
Agentic AI systems in accounting now automatically collect documents, extract data, validate completeness, flag exceptions, and produce review-ready work products. The profession is seeing a staffing shift: routine tasks require fewer hours at junior levels, creating demand for “digital seniors” who oversee AI workflows and communicate insights. Firms treating AI governance as daily operations rather than policy documents are pulling ahead in efficiency and client trust.

Mar 4, 2026 · IRC
Over two years, the International Rescue Committee deployed AI across refugee assistance, crisis response, and language support, reaching communities in low-bandwidth environments across dozens of languages. The IRC found that responsible AI in humanitarian settings requires operational infrastructure, not just ethical principles — including human override mechanisms and auditable outputs. Their Signpost platform uses AI for real-time translation and crisis anticipation. The work demonstrates AI augmenting frontline humanitarian capacity where there is no margin for error.

Mar 3, 2026 · Scientific American
Studies find AI-assisted developers ship more code but log longer hours and face more rollbacks, complicating the productivity narrative.

Mar 2, 2026 · Anthropic (Wiz customer story)
Wiz used Claude Code to complete a language migration estimated at two to three months of specialized engineering in roughly 20 hours, with 90% of engineers now using it daily.

Mar 2, 2026 · Anthropic (Rakuten customer story)
Rakuten achieved seven hours of sustained autonomous coding and a 79% reduction in time to market using Claude Code across its engineering organization.

Mar 2, 2026 · Anthropic (Ramp customer story)
Ramp deployed Claude Code across engineering, with 50% weekly active usage and up to 80% faster incident investigation through AI-built internal tooling.

Mar 2, 2026 · Anthropic (n8n customer story)
n8n built its AI Workflow Builder on Claude, cutting 80% of the manual work in assembling workflow automations and letting users go from idea to working workflow in minutes.

Mar 2, 2026 · Scott Cunningham
Economist Scott Cunningham generated a complete academic paper using Claude Code in under two hours for roughly $100: the AI originated the idea, applied a shift-share identification strategy, located data, ran the analysis, wrote the paper, addressed referee comments, and passed a code audit in two independent languages. Cunningham predicts AI-driven paper mills are coming at the researcher level. With research production trivially automated, the bottleneck shifts entirely to publication, and the credential system built around research faces structural pressure.

Mar 1, 2026 · OpenClaw
Anthropic’s open-source agent framework went from zero to 250K+ GitHub stars in under four months, overtaking React and Linux on the all-time leaderboard.

Mar 1, 2026 · Anthropic
Anthropic’s head of design Jenny Wen says the classic design process is “basically dead” as engineers with multiple Claude agents ship features faster than traditional research-and-mockup cycles allow.

Feb 28, 2026 · Amazon
Amazon has cut more than 30,000 workers since October, with CEO Andy Jassy aiming to run ‘like the world’s largest startup,’ redirecting capital from headcount to AI infrastructure. HR chief Beth Galetti called the cuts not ‘a new rhythm,’ but surviving workers report mounting workloads and managers pointing to ‘a bigger AI picture.’ Gartner analysts describe the approach as a deliberate playbook that rivals across tech and other sectors are watching closely.

Feb 27, 2026 · Zack Shapiro
A two-person boutique law firm handles workloads of much larger practices using Claude’s skills system, producing tracked-changes redlines, research memos, and real-time contract analysis without opening Word.

Feb 26, 2026 · Block
CEO Jack Dorsey lays off over 4,000 employees, nearly half the company, citing AI tools that enable smaller teams to fundamentally change how a company operates.

Feb 26, 2026 · Chris Blattman
A non-coding UChicago political economist built AI workflows with Claude Code in early 2026, cutting grant proposal drafting from 12 hours to 1 to 2 and automating inbox triage and project dashboards.

Feb 25, 2026 · Andrej Karpathy
Karpathy identifies December 2025 as the moment coding agents crossed a hard threshold; he built a full home video AI system in 30 minutes, a task that would have been a weekend project three months prior.

Feb 25, 2026 · Nvidia
Nvidia deployed OpenAI’s Codex to every engineer after CEO Jensen Huang mandated that every automatable task should use AI, making it one of the largest enterprise AI coding rollouts.

Feb 25, 2026 · Cursor
Cursor’s co-founder describes a shift from synchronous agents to autonomous cloud agents, with 35% of Cursor’s own PRs now created by agents in cloud VMs.

Feb 25, 2026 · Michael Truell (Cursor)
Cursor co-founder Michael Truell announced a third era of AI software development: agents running independently on cloud computers, tackling large tasks over longer timescales with minimal human direction. More than one-third of Cursor’s own merged PRs are now created by such agents. Truell predicts the vast majority of development work will shift to agents within a year. The developer’s role is evolving from writing code to managing a factory of AI agents.

Feb 24, 2026 · Aakash Gupta
Anthropic triggered five separate stock selloffs in four weeks; the iShares Software ETF fell 35% from peak, the sector’s worst month since 2008.

Feb 24, 2026 · Anthropic
Anthropic’s enterprise event positioned Claude as an orchestration layer atop existing software, easing investor fears and lifting battered software stocks.

Feb 24, 2026 · Harvard Business Review
HBR argues companies need a new “agent manager” role to orchestrate how AI agents learn, perform, and collaborate safely alongside human workers.

Feb 23, 2026 · Claude Code for Non-Coders
A developer explains that “100% AI-generated code” depends on a human-built context layer of specs, conventions, and architecture.

Feb 20, 2026 · METR
METR’s updated benchmarks show frontier AI agents’ autonomous task-completion time horizons continue growing exponentially, now handling multi-hour software tasks.

Feb 20, 2026 · Jason Gilmore
Craft-based programming teams are giving way to AI-powered factories where autonomous agents build from ideas and programmers manage systems.

Feb 20, 2026 · Andrej Karpathy
Karpathy explores a persistent orchestration layer sitting above LLM agents, enabling scheduling and long-running context with locally hosted personal AI.

Feb 19, 2026 · Mike Konczal
An economist describes integrating terminal AI tools like Claude Code into real-time economic data analysis: what used to require an hour of prep before 8:30 a.m. data releases now takes seconds, with robustness checks and chart generation handled autonomously. The judgment about what to measure remains human.

Feb 19, 2026 · IBM
While other companies cut headcount in favor of AI, IBM is tripling entry-level hiring across engineering, HR, and support, arguing that companies who hollow out junior pipelines now will struggle in five years.

Feb 19, 2026 · Ampcode
Ampcode retires its editor extensions because frontier models are capable enough that agent scaffolding is no longer the bottleneck; focus shifts to codebase structure.

Feb 19, 2026 · Stripe
Stripe’s deep dive into parallelizable cloud dev environments, isolated agent sandboxes, and engineering choices that make unattended agents safe at scale.

Feb 19, 2026 · Business Insider
Salesforce’s internal survey shows most employees feel AI increases their productivity, though fewer report reduced workloads, as the company bets its future on AI agents.

Feb 16, 2026 · Andrej Karpathy
LLMs are exceptionally good at code translation, and most software ever written will be rewritten many times over, potentially in new languages.

Feb 16, 2026 · Andrej Karpathy
LLMs are exceptionally good at code translation, and most software ever written will be rewritten many times over, potentially in new languages.

Feb 16, 2026 · Yascha Mounk
Political scientist Yascha Mounk argues that AI can now produce humanities scholarship indistinguishable from human academic writing, and that pretending otherwise is no longer credible. The essay confronts academics directly: the writing, analysis, and synthesis that define humanities work are now reproducible by AI at a quality that passes scrutiny. Mounk frames this as a fact already settled, not a projection. The question now is what follows.

Feb 15, 2026 · Business Insider
A six-person startup produced output that would have required 20 to 30 engineers five years ago; Canva’s CTO calls the agent-driven shift “really impressive.”

Feb 13, 2026 · OpenAI
An OpenAI team shipped a product with zero hand-written code using Codex, producing a million lines and 1,500 PRs in five months.

Feb 13, 2026 · OpenAI
OpenAI built a bespoke internal AI agent using GPT-5 and Codex to deliver analytical insights to teams in minutes rather than hours.

Feb 13, 2026 · Nikunj Kothari
A San Francisco field report where overnight agent runs, sober parties, and competitive agent counts have become the social currency of a restructuring workforce.

Feb 12, 2026 · Anthropic
Anthropic closed a $30B round at a $380B valuation, driven by $9B in run-rate revenue and Claude Code’s explosive enterprise adoption.

Feb 12, 2026 · Cloudflare
Cloudflare now converts web pages to Markdown automatically when AI agents request them, cutting token usage by up to 80%.

Feb 10, 2026 · Stripe
Stripe’s unattended coding agents produce over 1,000 human-reviewed pull requests per week with no human-written code, launched from a single Slack message.

Feb 8, 2026 · (Multiple)
Romance writer Coral Hart now publishes more than 200 novels a year with AI assistance, up from 10–12, earning six figures from roughly 50,000 sales. She uses Claude for prose, cycling through tools to work around content restrictions. The NYT profile documents a romance publishing industry being flooded by AI-assisted authors on Amazon while traditional publishers resist. For prolific self-publishers, AI has fundamentally changed the economics of creative output — and alarmed writers who built careers on craft.

Feb 7, 2026 · Simon Willison
StrongDM’s three-person team runs a “software factory” where agents write and test code with no human review, using Digital Twin clones.

Feb 5, 2026 · ICLE
A comprehensive ICLE review of the empirical AI productivity literature finds consistent gains of 15–50% in task completion time across writing, software development, customer support, accounting, law, and translation. Benefits are disproportionately large for less-experienced workers, compressing skill gaps within occupations. Despite rapid adoption — nearly 40% of U.S. adults used AI tools by late 2024 — aggregate labor market indicators through 2024–25 show limited disruption, mostly concentrated in entry-level segments of highly exposed occupations.

Jan 29, 2026 · Fortune
Boris Cherny ships 22 to 27 PRs daily with zero hand-written code; Anthropic is 70 to 90% AI-generated company-wide; an OpenAI researcher also writes none.

Jan 26, 2026 · Dario Amodei
Amodei frames powerful AI as humanity’s “technological adolescence,” calling for pragmatic, targeted interventions to manage risk without losing value.

Jan 22, 2026 · Alexander Kustov
Political scientist Alexander Kustov argues academia faces structural collapse from AI, not incremental change. AI can already produce publishable first-quartile social science research: Tibor Rutar generated a full paper from prompts alone, Paul Novosad replicated the feat in 2–3 hours. Kustov’s ten theses hold that the research production bottleneck is gone, the publishing system isn’t built for this volume, and most academics — the most dispositionally conservative institution on the planet — haven’t noticed the ground shifting.

Jan 15, 2026 · McKinsey
McKinsey announced a workforce of 20,000 AI agents and embedded an AI competency test into its graduate hiring process. The agents handle knowledge synthesis and research firm-wide. Requiring AI aptitude in new hires marks a shift in what the firm is selecting for: the ability to direct and work alongside agents is now part of the baseline expected of incoming consultants, alongside traditional analytical and interpersonal skills.

Jan 14, 2026 · Sequoia Capital
Sequoia declares AGI has functionally arrived through long-horizon coding agents that can work autonomously for hours, iterating through problems like a human.

Jan 13, 2026 · One Useful Thing
Ethan Mollick demonstrates Claude Code autonomously building a complete startup website in over an hour, calling it a major leap in sustained autonomous AI work.

Jan 6, 2026 · Tailwind Labs
AI caused docs traffic to drop 40% and revenue nearly 80%, forcing layoffs of 75% of the engineering team despite continued popularity growth.

Jan 2, 2026 · Erik Brynjolfsson
Stanford Digital Economy Lab director Erik Brynjolfsson argues 2025 planted the seeds of a profound transformation, with AI agents now handling execution while humans focus on problem definition and evaluation. A PwC survey found 79% of companies leveraging agentic AI. Brynjolfsson frames the shift through his research on tacit knowledge: AI has learned to execute complex action sequences without explicit rules, finally automating what was considered irreducibly human.

Dec 19, 2025 · PwC
PwC uses Harvey AI to automate legal entity reviews and liquidation workflows, shifting M&A support from manual checklists to end-to-end AI-driven execution across global projects.

Dec 3, 2025 · Anthropic
Anthropic announced that Claude Code crossed $1 billion in revenue and simultaneously acquired Bun, the fast JavaScript runtime built by Jarred Sumner. Bun’s toolchain joins the Claude Code ecosystem, accelerating infrastructure for agentic development workflows. The $1 billion milestone arrived in roughly two years. That is a measure of how fast enterprise developers moved from evaluating AI coding tools to depending on them for core engineering work.

Nov 6, 2025 · (Multiple)
By mid-2025, the FDA had cleared 873 AI algorithms for medical imaging — 115 added that year alone — making radiology the single largest target for AI among clinical specialties. Tools now flag potential cancers in mammograms and lung scans, triage urgent cases, and draft preliminary reports. Adoption is growing unevenly across institutions. Foundation models capable of linking images with clinical text are entering early deployment.

Oct 30, 2025 · Brex
A content designer with no coding background now builds Figma plugins and ships features using Claude Code, illustrating how AI extends engineering capability to non-engineers.

Aug 18, 2025 · MIT Sloan Management Review
AI coding tools can make developers up to 55% more productive, but MIT researchers warn that rapid deployment creates dangerous technical debt, especially in legacy systems. AI-generated code can destabilize architecture in ways that only surface months later, potentially erasing productivity gains and crippling scalability.

Aug 13, 2025 · Goldman Sachs
Over 50% of Goldman’s 46,000 employees adopted the internal GS AI Platform, yielding 20% productivity gains for coders and a 15% drop in post-release bugs, with CEO targeting 100% adoption by 2026.

Jul 24, 2025 · Walmart
Walmart’s CTO describes four “super agents” unifying customer, associate, partner, and developer experiences across the company on a shared agentic AI framework.

Apr 29, 2025 · CNBC
Nadella revealed 20 to 30% of Microsoft’s code is AI-generated; Zuckerberg predicted Meta will reach 50% within a year.

Apr 7, 2025 · Salesforce
Developers are moving up the stack to system design and agent management, with CodeGenie handling 7 million lines and saving 30,000 hours monthly.

Apr 1, 2025 · Stanford HAI
Stanford’s annual AI report finds inference costs dropped 280-fold since 2022, hardware costs fell 30% annually, and open-weight models closed the gap with closed models to just 1.7%.

Feb 4, 2025 · (Multiple)
A randomized study of 5,172 customer support agents published in the Quarterly Journal of Economics found AI assistance raised productivity by 15% on average, measured by issues resolved per hour. Gains were largest for lower-skilled and newer workers, who improved both speed and quality; top performers saw small speed gains with slight quality declines. AI also facilitated worker learning, improved English fluency among international agents, and made customers measurably more polite.

Nov 1, 2024 · Google
CEO Sundar Pichai revealed AI generates over 25% of all new code at Google, with engineers reviewing and accepting AI contributions.

Feb 27, 2024 · Klarna
Klarna’s OpenAI-powered assistant handled 2.3 million conversations in its first month, doing the work of 700 full-time agents with equal satisfaction scores and an estimated $40 million profit improvement.

Aug 2, 2023 · McKinsey
McKinsey’s generative AI platform Lilli reached 72% firm-wide adoption, with over 500,000 prompts monthly and colleagues reporting up to 30% time savings in searching and synthesizing knowledge.

Jul 13, 2023 · (Multiple)
A preregistered experiment recruited college-educated professionals to complete incentivized writing tasks with and without ChatGPT. Workers assigned AI finished faster, produced higher-quality output, and reported enjoying the work more. Workers with weaker baseline skills benefited most. Published in Science, the Noy and Zhang study was among the first rigorous demonstrations that generative AI could raise skilled professional productivity — and that gains would flow unequally, compressing skill differences within occupations.

We are diving in. Will AI kill us all? Some of us? Now? Later? Debates about artificial intelligence can stall when we hold different assumptions about the nature of the risk and different moral intuitions about how to respond to that risk. In this framework I try to clarify those disagreements by separating two distinct questions that get conflated when I talk to others about AI:

1. What kind of technology is superintelligent AI?

2. How should we respond to that risk?

We can think of each question as a model. The first model maps how us understand superintelligence and its risk along two axes: whether it is continuous with past technologies or represents a genuine discontinuity, then whether its impacts are manageable or unmanageable. The second model examines our moral action preferences along two axes: whether we are more risk averse or more risk accepting and whether we favor preservation or aspiration.

Taken together, these two models offer a clearer picture of the risk landscape by separating what we believe about superintelligence from how we are motivated to act in response. I developed this framework to clarify my own thinking, not to model everything important about AI risk, AI safety, and certainly not to tell you where to land.

Roadmap for this article:

Model One - A model of superintelligence risk
Model Two - A model of moral action
Combined - The 16 risk and moral action pathways
Beyond my framework

Model One - A model of superintelligence risk

The first model helps us decide what kind of world we are in. This is descriptive, not prescriptive. One way to think about the risk of superintelligence in our world is to consider two axes together.

Continuity ←→ Discontinuity: is superintelligence fundamentally continuous with human technology or a new phenomenon not analogous to any prior human experience?

Optimism ←→ Pessimism: is coexistence with superintelligence manageable or unmanageable?

Continuity Position:

Risk from superintelligence looks like other powerful and complex technologies we coexist with, such as the internet and nuclear weapons. The risk is serious but governable with institutions, incentives, and iteration. The main danger is how humans deploy, govern, and incentivize AI systems.

Discontinuity Position:

Superintelligence will be a break from past technologies and will significantly surpass human intelligence and capability. Mistakes will be irreversible. The main danger is what superintelligence will do with humans once it surpasses us.

What is each outlook trying to protect in the face of risk?

Whether we think superintelligence will be continuous or discontinuous with human evolution so far, we likely have either an optimistic or pessimistic view of expected outcomes. Optimism and pessimism here refer to expectations about manageability, not emotional outlook. While understanding that most people do not rigidly occupy a single grid space, let’s explore the possible combinations.

Continuity + Optimism = The Builders

This pair understands superintelligence to be primarily a tool for improving lives. If superintelligence is a normal technology, the primary risk is unnecessary restraint and lost upside.

Superintelligence is a powerful technology we can deploy and manage. In this view, the greatest risks are: missed opportunities for broad human benefit, overregulation that slows beneficial progress, and treating superintelligence as categorically different from prior technologies without sufficient justification.

Examples:

• AI as Normal Technology, by Arvind Narayanan & Sayash Kapoor

• A Guide to Understanding AI as Normal Technology, by the same

Continuity + Pessimism = The Institutionalists

This pair sees how powerful technologies go wrong in the hands of powerful humans. The primary risk isn’t innovation; rather, it’s weak governance, bad incentives, and a lack of accountability that enable humans to use superintelligence for irreversible harm.

Danger comes from human governance failures. In this view, the greatest risks are: institutional failure and regulatory lag, concentration of power and its misuse at scale, and familiar human incentives that repeat old harms.

Examples:

• Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity, by Daron Acemoglu & Simon Johnson

• Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence, by Kate Crawford

Discontinuity + Optimism = The Navigators

This pair sees superintelligence as a phase change, so early errors matter disproportionately and can still be avoided. We still owe future generations the benefits of progress if we can get this right.

Phase change requires careful steering. In this view, the greatest risks are: early design mistakes that scale irreversibly, lack of coordination at critical inflection points, and failure to invest in alignment and safety early.

Examples:

• AI Impacts conversation, by Paul Christiano

• Human Compatible: Artificial Intelligence and the Problem of Control, by Stuart J. Russell

Discontinuity + Pessimism = The Existentialists

This pair sees superintelligence as a phase change and that a single mistake could permanently end our ability to choose our future. Even rare errors must take precedence over all other moral goals, even at the cost of slowing progress.

Irreversible stakes and asymmetric risk dominate. In this view, the greatest risks are: irreversible loss of human control, single-point failures with no recovery path, and overconfidence in our ability to correct mistakes later.

Example: Is Power-Seeking AI an Existential Risk?, by Joseph Carlsmith

Model Two - A model of moral motivation and action

The first model helped us decide what we believe about the world with superintelligence in it. The next model helps explore how we can act in response to what we believe about that world — that is, how our moral intuitions shape decisions when certainty is out of reach. I prioritize two axes for consideration, though there are others.

Risk-Averse ←→ Risk-Accepting: Do we treat risk as a cost to be minimized before acting, or as a condition we accept in order to act on what we value?

Preservation ←→ Aspiration: Are we primarily motivated by protecting what exists, or by reaching for what could be?

This model does not rank the options or endorse any one stance. As in the first model, most people do not rigidly occupy a single grid space.

For each quadrant, I identify both the ultimate end and the primary virtue (the disposition required to pursue it).

Avoid Risk + Preserve. The highest good is survival and preservation of existing value. The highest virtue is caution, even in the face of irreversible loss. This stance prescribes avoiding actions that could permanently eliminate our ability to choose. Restraint is morally required when the stakes are existential, even if it means accepting a diminished or constrained future for humans.

Avoid Risk + Aspire. The highest good is gradual improvement of human well-being. The highest virtue is prudence. Make progress with guardrails, reversibility, and humility. This stance prescribes advancing only where mistakes can be corrected. Progress is moral only when it remains under control, even if it means a slower or more modest trajectory for humans.

Accept Risk + Preserve. The highest good is the long-term survival of humanity as we know it now, even under threat. The highest virtue is the vigilance, the willingness to act decisively, but responsibly, to protect who we are. This stance prescribes taking necessary risks to defend humanity’s present. Inaction can itself be a fatal choice.

Accept Risk + Aspire. The highest good is a future defined by striving, discovery, and transcendence. The highest virtues are inseparable here, curiosity and the courage to act decisively in pursuit of new ways of being, even in the face of uncertainty. This stance prescribes pushing forward despite existential danger, because mere survival without striving, discovery, or transcendence would be the most tragic kind of loss.

Combined - The 16 risk and moral action pathways

When each of these action models are paired with an AI risk model, we can begin to see how moral action pathways vary depending on what we believe about a world with superintelligence. Let’s look at all the combinations.

The Builders share belief in continuity + optimism

Superintelligence is not a phase change, and it’s manageable. Pathways:

1. Avoid Risk + Preserve: The Cautious Builder banks good odds but hedges anyway. Even if superintelligence is manageable, we should move cautiously. Past technologies brought unforeseen harms despite optimistic projections. Preserving optionality matters even when the odds look good.

2. Avoid Risk + Aspire: The Steady Builder trusts iteration and correction. Superintelligence is a tool like any other. Improve it steadily, within reasonable bounds, and the future will be better than the present. No need for dramatic action in either direction.

3. Accept Risk + Preserve: The Vigilant Builder sees continuous risk as a continuous threat, but one that is manageable with vigilance. Even tractable technologies can be weaponized by bad actors, and history shows that beneficial technologies require active defense, not just optimism. If superintelligence is continuous with the past, the danger lies not in the technology itself but in how humans choose to use it — which means the defense is also in human hands.

4. Accept Risk + Aspire: The Bold Builder treats hesitation as the real risk. If superintelligence is manageable and continuous with past technologies, the moral imperative is to push forward. Slowing down wastes human potential and delays benefits that could reduce suffering now.

The Institutionalists share belief in continuity + pessimism

Superintelligence is not a phase change, but we’ll struggle to manage it. Pathways:

1. Avoid Risk + Preserve: The Regulatory Institutionalist prioritizes institutional safeguards over speed. If superintelligence amplifies familiar human failures like greed, authoritarianism, and capture, then the answer is restraint, regulation, and guardrails before deployment. Don’t hand new weapons to the same bad actors.

2. Avoid Risk + Aspire: The Reformist Institutionalist believes better systems unlock better technology. If superintelligence’s danger comes from human institutional failures rather than the technology itself, then the path to a better future runs through reforming those institutions. We should improve governance, accountability, and incentive structures while advancing AI capabilities only where those reforms have taken hold.

3. Accept Risk + Preserve: The Defensive Institutionalist fights fire with fire, but the fire is systemic. The same institutional failures that make superintelligence dangerous, such as regulatory capture, concentration of power, and weak accountability, also make waiting dangerous. This stance prescribes building countervailing power before those failures lock in, not just defending against individual misuse. The threat is that the systems meant to govern the technology are already compromised.

4. Accept Risk + Aspire: The Competitive Institutionalist races toward good outcomes before bad actors win. If superintelligence amplifies existing human failures, then hesitation only ensures bad actors move first. The path to human flourishing may require racing to build beneficial AI systems before malicious ones dominate, accepting the risk of our own mistakes to prevent worse outcomes from others.

The Navigators share belief in discontinuity + optimism

Superintelligence is a phase change, but we can manage it. Pathways:

1. Avoid Risk + Preserve: The Cautious Navigator protects the present before reaching for the future. If superintelligence is a genuine phase change but still solvable, the first priority is ensuring we don’t lose what we have. Before pursuing transformative benefits, we must secure humanity’s survival and autonomy. Caution and preservation come before aspiration when the stakes stakes admit no second chances.

2. Avoid Risk + Aspire: The Careful Navigator invests heavily in safety before capability, from a position of genuine optimism. If superintelligence is a genuine phase change but still solvable, careful progress is the winning path, not merely the cautious one. This stance prioritizes alignment before scaling because getting it right is achievable, and we owe future generations both the benefits of superintelligence and the wisdom required to deliver them responsibly.

3. Accept Risk + Preserve: The Strategic Navigator accepts risk now as a strategic investment in a shapeable outcome. If superintelligence is a phase change but still solvable, the window to act may be brief. Acting decisively now is how we keep the problem solvable. This stance prescribes bold, high-stakes action not as a last resort, but as a forward-looking bet that the right intervention at the right moment can preserve humanity's ability to choose its future.

4. Accept Risk + Aspire: The Bold Navigator: This position bets on transformation from a position of genuine optimism. If superintelligence is a phase change with transformative upside and the problem is still solvable, boldness is a calculated bet on a shapeable future rather than reckless action. The potential for transcendence is real and accessible if we act decisively. A future of radical improvement is achievable and requires radical commitment.

The Existentialists share belief in discontinuity + pessimism

Superintelligence is a phase change, and we are unlikely to manage it well. Pathways:

1. Avoid Risk + Preserve: The Quiet Existentialist treats restraint as the highest form of responsibility. If superintelligence is a phase change with a high likelihood of irreversible catastrophe, then the moral imperative is maximum restraint. Any action that could trigger loss of human control must be avoided, even if it means forgoing potential benefits. When a single mistake could end everything, caution is the only responsible path.

2. Avoid Risk + Aspire: The Careful Existentialist advances carefully, but without confidence that care will be enough. If superintelligence poses unprecedented risk and is unlikely to be fully managed, the moral case for progress doesn't disappear but becomes tragic rather than hopeful. This stance prescribes building only what can be controlled, testing before deploying, and halting if alignment fails. It moves forward not as an optimistic bet but as a moral obligation pursued against the odds.

3. Accept Risk + Preserve: The Defensive Existentialist accepts dangerous action as the price of survival, without optimism about what that action can achieve. If superintelligence poses existential risk that cannot be fully managed, waiting becomes its own form of surrender. This stance prescribes preemptive action, disrupting dangerous trajectories, and taking moves that themselves carry risk, not because a good outcome is likely, but because passive inaction forfeits even the chance of one. The action here is desperate, not strategic.

4. Accept Risk + Aspire: The Pathbreaking Existentialist accepts existential stakes as the price of a transformative future, even if catastrophe awaits. If superintelligence entails irreversible catastrophe, then a future defined only by survival of the present — without striving, discovery, or transcendence — would itself be a kind of loss. This is the most uncomfortable position in my framework. Press forward despite believing the odds are against us, because a future of mere survival without transformation is seen as a failure in itself. Risk is less tragic than the idling of human progress.

Hidden Variables

The two models above capture the disagreements I think about the most. They don’t capture all the disagreements. Other dimensions, such as timeline, concentration of control, epistemic confidence, and more, cut across the 16 pathways and produce divergent strategies even among people who share the same pathway.

Timeline, or takeoff speed, is perhaps the most important variable not explicitly in my two models. Two people can agree completely about the nature of the risk and the moral response, yet prescribe opposite actions because they hold different assumptions about how quickly superintelligence will arrive.

I didn’t include takeoff speed explicitly because I think its implications are easy to surface. In my framework, I wanted to surface the less obvious dimensions of risk avoidance versus acceptance and preservation versus aspiration. Once we understand where we are on those dimensions, it’s not too hard to adjust our action preferences based on how soon we think superintelligence will be broadly active.

Let’s walk through two examples.

The Cautious Navigator under fast and slow assumptions

Consider the Cautious Navigator: Discontinuity + Optimism, Avoid Risk + Preserve. This position holds that superintelligence is a genuine phase change, that the problem is still solvable, and that the first priority is securing what we have before reaching for what we could gain. Two Cautious Navigators who agree on all of this still take different actions depending on their assumptions about how soon superintelligence will be broadly active.

Under a slow takeoff assumption — superintelligence emerges over decades, with visible milestones and warning signs along the way — the Cautious Navigator has time to build. They will prefer patient, institutional work, like investing in alignment research, developing international governance frameworks, building monitoring infrastructure, and iterating on safety standards as capabilities grow. Caution looks like careful preparation. The slow timeline makes preservation compatible with measured, deliberate action.

Under a fast takeoff assumption — superintelligence could emerge in 18 months with little warning — the same Cautious Navigator faces a radically compressed decision space. The same moral commitment to preservation now demands urgency rather than patience: immediate moratoria on the most dangerous research, emergency coordination between labs and governments, or hard capability limits imposed before governance frameworks are ready. Caution manifests as an emergency action because, in this view, preservation is incompatible with waiting.

The Bold Navigator under fast and slow assumptions

Consider the Bold Navigator: Discontinuity + Optimism, Accept Risk + Aspire. This position holds that superintelligence is a genuine phase change, that the problem is still solvable, and potential for radical human flourishing justifies accepting significant risk. Two Bold Navigators who agree on all of this pursue very different strategies depending on their assumptions about how soon superintelligence will be broadly active.

Under the assumption of a slow takeoff, the Bold Navigator has room to be ambitious and strategic. They will invest heavily in frontier research, push for open collaboration between labs and the government, fund moonshot alignment approaches that might take years to mature, and advocate for policies that accelerate beneficial capabilities rather than restrict them. Boldness looks like sustained, visionary investment. The slow timeline lets aspiration unfold through compounding gains, and the Bold Navigator treats each milestone as an opportunity to raise the ceiling rather than to consolidate.

Under a fast takeoff assumption, the same Bold Navigator faces a closing window. The same moral commitment to transformation now demands concentrated, high-stakes bets rather than broad portfolio plays. Their preference shifts toward backing a single leading alignment-capable lab, deploying partially validated systems to gain real-world feedback before it’s too late. Boldness in this timeline is a decisive commitment under radical uncertainty. The fast timeline makes aspiration incompatible with waiting for better evidence.

Where the Cautious Navigator’s fast takeoff response is to hit the brakes, the Bold Navigator’s is to steer harder. Both share the Navigator’s belief that the phase change is real and solvable, but their moral orientations produce opposite instincts as the clock runs out. One secures what exists, the other reaches for what could be.

Timeline assumptions reshape the logic of all 16 pathways in the framework. I invite you to apply this lens to any of the other 14 pathways and notice how the recommended action shifts. Sometimes subtly, sometimes dramatically, depending on whether superintelligence is decades away or imminent.

Not all missing dimensions work this way. Some perspectives don’t add axes to the framework, but instead challenge its premises entirely.

Lenses That Resist My Classifications

My two models above assume we can locate ourselves on their axes, that we hold some view about the nature of superintelligence and some view about how to act under uncertainty. Some perspectives reject the framing itself. The following lenses don’t fit neatly into either of my models. They challenge the premises rather than choosing quadrants. These views are not errors. I include them to acknowledge that my framework has boundaries and doesn’t capture every serious position.

The Evolutionary Lens sidelines moral intent almost entirely. In this view, the future of AI will be shaped less by what people believe is right and more by market competition, politics, and arms race dynamics. The systems that win may not be the ones we morally prefer. Moral frameworks like mine might function as post-hoc rationalizations of positions people were going to hold anyway due to their structural incentives.

The Virtue Lens refuses to optimize anything directly. In this view, who we become matters more than what we build. The moral risk isn’t just what superintelligence does to humans, but who we become by building it. The second of my two models frames moral action in terms of outcomes (preserve vs. aspire, avoid vs. accept risk), and virtue ethics is fundamentally about character, not outcomes. This lens is rejecting the consequentialist scaffolding underneath that model.

The Post-Human Lens refuses human supremacy. In this view, human extinction is not the worst outcome. Post-human futures could be superior. Much AI debate assumes human continuation is the ultimate good, but that assumption is up for grabs in the post-human lens. Extinction is less catastrophe and more transition.

Final Thought

I offer these models in case they are helpful. For me, they are useful for the questions they surface, not as a final word on what superintelligence demands of us or what we ought to demand of each other.

A Mumbai HR software company placed an autonomous AI named Ivy into a live Junior HR Executive role — real Slack access, real company email, real KPIs — and ran the experiment for seven days. In the same week, an NBER paper projected 502,000 AI-driven job cuts in 2026, nine times last year’s figure, alongside Oracle’s elimination of 10,000 engineers and technical specialists, no performance explanation given. Software engineer and Django co-creator Simon Willison described AI’s impact on engineering as a structural shift in an April 2026 podcast appearance.

This week’s update also reaches back to fill a gap in the log: by mid-2025, the FDA had cleared 873 AI algorithms for medical imaging, making radiology the single largest clinical AI target. A separate randomized trial of 515 startups found that helping firms discover where to deploy AI produced 1.9x higher revenue and 12% more tasks completed.

Apr 3, 2026 · Asanify
Asanify placed an autonomous AI named Ivy into a live Junior HR Executive role with Slack access, a company email, HR admin credentials, and KPIs matching those of a human hire. Ivy handled real employees with real consequences over seven days. The company framed it as “eating its own cooking” — they build AI-native HR software and needed to know whether their technology could genuinely perform an entry-level role, not merely assist one.

Apr 2, 2026 · Simon Willison
Software engineer and Django co-creator Simon Willison described AI’s impact on engineering as a structural shift in an April 2026 podcast appearance. He called November 2025 an inflection point, named software engineers as bellwethers for all knowledge workers, and said AI has broken his ability to estimate project timelines. He flagged dark factories, fully AI-run operations, as an emerging model, and observed the transition is especially hard for mid-career engineers.

Mar 31, 2026 · Oracle
Oracle let go roughly 10,000 employees in late March 2026, eliminating roles for senior engineers, architects, operations leaders, program managers, and technical specialists. The company declined to comment on the cause, but executives had previously stated that AI tools “enabled fewer employees to do more work.” The cuts were described as not performance-based, signaling a structural workforce reduction tied to the company’s heavy AI investment.

Mar 30, 2026 · INSEAD/HBS
A randomized controlled trial across 515 global startups found that helping firms discover where to deploy AI produced dramatic firm-level gains: treated startups completed 12% more tasks, were 18 percentage points more likely to acquire paying customers, and generated 1.9x higher revenue, while reducing their need for external capital by 39.5%. Each additional AI use case discovered added ~26% more revenue.

Mar 27, 2026 · Tesla
Tesla this year phased out car models, including a popular luxury sedan, to stand up a new Optimus humanoid production line. Elon Musk, who frames the project as building a world of abundance where work is optional, merged SpaceX’s AI unit with Tesla’s robot program. At least three other major tech investors are funding competing humanoid programs. The push represents a direct bet that physical AI, not just software, will make human labor obsolete.

Mar 26, 2026 · (Multiple)
An NBER working paper drawing on 750 CFOs found AI-driven job cuts in 2026 are projected at 502,000 — nine times the 55,000 cut in 2025. ADP Research, surveying 39,000 workers across 36 countries, found only 22% strongly agreed their job was safe. A counterintuitive finding: workers using AI tools report time on some tasks has increased up to 346%, as adoption without workflow redesign creates friction rather than speed.

Nov 6, 2025 · (Multiple)
By mid-2025, the FDA had cleared 873 AI algorithms for medical imaging — 115 added that year alone — making radiology the single largest target for AI among clinical specialties. Tools now flag potential cancers in mammograms and lung scans, triage urgent cases, and draft preliminary reports. Adoption is growing unevenly across institutions. Foundation models capable of linking images with clinical text are entering early deployment.

Read the full blog:

Welcome to the Treeline! The view opens up here.

This update adds new entries to my ongoing log of AI’s transformation of work, the good and the bad. It also includes a couple of older articles filling gaps in the log.

Forty-nine percent of jobs now have at least a quarter of their tasks performed using Claude, according to Anthropic’s March Economic Index, with usage broadening into lower-wage occupations. Experienced users attempt harder tasks and succeed more often, creating a self-reinforcing advantage. The skill that compounds fastest right now is learning to work with AI itself.

Warp stopped buying SaaS entirely, replacing subscriptions with agents and just-in-time apps built from raw APIs. Basecamp opened its whole platform to agent access the same week. Google engineer Addy Osmani documented the convergence across Cursor, GitHub Copilot, and Claude Code Web: developers now specify intent and review diffs rather than write code directly. Armin Ronacher, creator of Flask, observed that time saved gets immediately captured by competition, and the pressure to ship at inference speed is eroding the commitments that durable software depends on.

This week’s log entries

Mar 27, 2026 · Zach Lloyd (Warp)
Warp CEO Zach Lloyd details how his company stopped buying SaaS, replacing subscriptions with agents, skills, and just-in-time apps. They migrated 266 documentation pages from a hosted platform to Markdown in hours, replaced community-insight SaaS with a monitoring agent that suggests replies and self-improves, made data analytics self-serve via agent skills with BigQuery access, and built custom recruiting tools from raw APIs. Savings exceeded $10k per year on docs alone.

Mar 25, 2026 · Basecamp
Basecamp announced full CLI and agent access to its platform, letting AI agents do anything a human can — write documents, manage to-dos, answer check-ins. Claude, Codex, OpenCode, and Cursor can all connect directly. The move reframes project management software as agent-native infrastructure, positioning the user’s AI tooling at the center rather than the app’s own interface.

Mar 24, 2026 · Anthropic
Anthropic’s latest Economic Index finds that 49% of jobs now have at least a quarter of their tasks performed using Claude, up from a narrower set of high-value coding tasks in late 2025. Usage is diversifying toward lower-wage occupations as adoption broadens. The report’s central finding: experienced users attempt higher-value tasks and elicit more successful responses, suggesting early AI adoption creates self-reinforcing skill advantages that could widen inequality between early and late adopters.

Mar 21, 2026 · Addy Osmani
Google engineer Addy Osmani documents a structural shift in developer tools: the primary work surface is moving from the code editor to an agent orchestration layer. Cursor’s new Glass interface, GitHub Copilot Agents, Conductor, and Claude Code Web all converge on the same pattern — developers specify intent, delegate to parallel agents running in isolated workspaces, then review diffs. The agent, not the file, is becoming the unit of work.

Mar 21, 2026 · Azeem Azhar (Exponential View)
Azeem Azhar analyzes Jensen Huang’s GTC declaration that every company needs an OpenClaw strategy. The core argument: the AI debate fixates on capability when diffusion is what matters. The harness — not the engine — is the revolution, just as the automobile mattered more than the combustion engine alone. Tokens are now a productive input as fundamental to knowledge work as electricity. Organizations treating token budgets as an IT cost center rather than a business input are dangerously behind.

Mar 20, 2026 · Armin Ronacher
Armin Ronacher, creator of Flask and a developer at the center of AI tooling, reports that time saved by AI gets immediately captured by competition: anyone who pauses is outmaneuvered by someone filling every freed-up hour with new output. A YC batch last year saw startups vanish without notifying customers, treating proper shutdowns as wasted time. Ronacher argues the pressure to ship at inference speed is eroding the trust, community, and long-term commitment that durable software depends on.

Mar 5, 2026 · Texas Society of CPAs
Agentic AI systems in accounting now automatically collect documents, extract data, validate completeness, flag exceptions, and produce review-ready work products. The profession is seeing a staffing shift: routine tasks require fewer hours at junior levels, creating demand for “digital seniors” who oversee AI workflows and communicate insights. Firms treating AI governance as daily operations rather than policy documents are pulling ahead in efficiency and client trust.

Subscribe now

Read the full log at

This week’s update adds 21 entries to the full log!

A handful are from the past few days and a larger group of older pieces backfilled to close gaps in the log. The new entries land on a week where Hex reported that AI agents created more cells than humans for the first time, OpenAI acquired the company behind Python’s most-used development tools, and Amazon called an emergency all-hands after AI-assisted code changes took the site down for six hours. Atlassian cut 1,600 people, citing AI. The layoff announcements keep arriving in AI’s language, though analysts increasingly note that the language might be doing more work than the technology.

The backfilled entries surface a pattern worth sitting with. A BCG experiment found AI users completed more tasks, faster, with higher quality, except on problems just outside AI’s capability frontier, where they performed measurably worse. Two separate studies in the QJE and Science confirmed that junior workers gain the most from AI assistance, compressing skill gaps within occupations. And three academics arrived independently at the same conclusion: AI can now produce publishable humanities and social science research.

Subscribe now

Mar 20, 2026 · (Multiple)
Q1 2026 has already seen more tech layoffs than Q1 2025, with Amazon, Meta, Salesforce, and Block among companies citing AI. But analysts are calling it: much is standard cost-cutting dressed in AI language. Block’s Jack Dorsey cut nearly half the workforce while ‘embracing AI,’ prompting critics to call the framing ‘convenient.’ The Federal Reserve Bank of St. Louis found a striking correlation between AI prevalence and unemployment increases since 2022, but a Harvard Business Review report suggests many layoffs are driven by AI’s potential, not its performance.

Mar 19, 2026 · Hex
On March 10, 2026, agents created more cells in Hex than humans did for the first time. Hex CEO Barry McCardel announced the milestone on LinkedIn: in a product where queries, charts, and code are the atomic units of work, the balance tipped from human-created to agent-created. Usage has gone parabolic, McCardel wrote, and the transition happened faster than expected as agentic surfaces became deeply embedded in core workflows.

Mar 19, 2026 · Astral
OpenAI is acquiring Astral, the company behind Python tools Ruff, uv, and ty, collectively reaching hundreds of millions of downloads monthly. Astral will join OpenAI’s Codex team. CEO Charlie Marsh framed the move as working at ‘the highest-leverage thing we can do’ at the frontier of AI and software. The acquisition consolidates foundational Python development tooling inside the AI company now building coding agents intended to automate engineering work.

Mar 17, 2026 · Intercom
Intercom engineer Brian Scanlan detailed how the company built a full-stack engineering platform on Claude Code: 13 plugins, 100-plus skills, and automated hooks. The most striking component is a read-only Rails production console via MCP, giving Claude live access to execute arbitrary Ruby against production data for feature flag checks, cache inspection, and business logic validation, with Okta authentication and blocked critical tables as safety gates.

Mar 12, 2026 · (Multiple)
An extended NYT Magazine portrait of software engineers who barely write code anymore. Developers now spend their days directing AI agents in plain English. At Hyperspell, a task that once took a full day takes 30 minutes. Prompt files have become behavioral contracts — including stern commands like ‘pushing code that fails pytest is unacceptable and embarrassing’ to shape AI output. The piece documents a shift already underway: engineer as manager.

Mar 11, 2026 · Atlassian
10% of Workforce — Citing AI* — Atlassian cut 1,600 roles, 10 percent of its global workforce, in response to AI’s effect on how software is built. Co-founder and CEO Mike Cannon-Brookes framed it as choosing to adapt rather than a cost-cutting emergency. Australia’s largest listed technology company joins a list of enterprise software firms shedding staff as AI tools compress engineering output, shrinking the headcount needed to sustain development at scale.

Mar 11, 2026 · Boston Consulting Group
In a preregistered experiment with 758 BCG knowledge workers, AI assistance completed 12.2% more tasks, 25.1% faster, with higher quality — on tasks within AI’s capability frontier. But for a complex managerial task outside that frontier, AI users were 19% less likely to produce correct solutions. The HBS/BCG study coined ‘jagged technology frontier’ to describe how AI’s impact varies unpredictably even across seemingly similar tasks in the same knowledge workflow.

Mar 10, 2026 · Amazon
Amazon’s ecommerce leadership called an emergency all-hands after a ‘trend of incidents’ linked to ‘Gen-AI assisted changes’ — including a six-hour site outage that left customers unable to complete transactions. A company memo cited ‘novel GenAI usage for which best practices and safeguards are not yet fully established.’ Junior and mid-level engineers now require senior sign-off on AI-assisted changes. Amazon Web Services suffered separate AI-related incidents in the same period.

Mar 5, 2026 · Cursor
and Questions Its Own Purpose* — Forbes profiles Cursor, the fastest-growing AI coding company, as it confronts an uncomfortable question: if AI agents can write code autonomously without a human code editor, what is a code editor for? Cursor grew explosively as a developer tool but now faces pressure from pure agent products that bypass editors entirely. The company is building deeper agentic capabilities, competing in a landscape its own success helped create.

Mar 4, 2026 · IRC
Over two years, the International Rescue Committee deployed AI across refugee assistance, crisis response, and language support, reaching communities in low-bandwidth environments across dozens of languages. The IRC found that responsible AI in humanitarian settings requires operational infrastructure, not just ethical principles — including human override mechanisms and auditable outputs. Their Signpost platform uses AI for real-time translation and crisis anticipation. The work demonstrates AI augmenting frontline humanitarian capacity where there is no margin for error.

Mar 2, 2026 · Scott Cunningham
Economist Scott Cunningham generated a complete academic paper using Claude Code in under two hours for roughly $100: the AI originated the idea, applied a shift-share identification strategy, located data, ran the analysis, wrote the paper, addressed referee comments, and passed a code audit in two independent languages. Cunningham predicts AI-driven paper mills are coming at the researcher level. With research production trivially automated, the bottleneck shifts entirely to publication — and the credential system built around research faces structural pressure.

Feb 28, 2026 · Amazon
Amazon has cut more than 30,000 workers since October, with CEO Andy Jassy aiming to run ‘like the world’s largest startup’ — redirecting capital from headcount to AI infrastructure. HR chief Beth Galetti called the cuts not ‘a new rhythm,’ but surviving workers report mounting workloads and managers pointing to ‘a bigger AI picture.’ Gartner analysts describe the approach as a deliberate playbook that rivals across tech and other sectors are watching closely.

Feb 16, 2026 · Yascha Mounk
Political scientist Yascha Mounk argues that AI can now produce humanities scholarship indistinguishable from human academic writing, and that pretending otherwise is no longer credible. The essay confronts academics directly: the writing, analysis, and synthesis that define humanities work are now reproducible by AI at a quality that passes scrutiny. Mounk frames this as a fact already settled, not a projection. The question now is what follows.

Feb 8, 2026 · (Multiple)
Romance writer Coral Hart now publishes more than 200 novels a year with AI assistance, up from 10–12, earning six figures from roughly 50,000 sales. She uses Claude for prose, cycling through tools to work around content restrictions. The NYT profile documents a romance publishing industry being flooded by AI-assisted authors on Amazon while traditional publishers resist. For prolific self-publishers, AI has fundamentally changed the economics of creative output — and alarmed writers who built careers on craft.

Feb 5, 2026 · ICLE
A comprehensive ICLE review of the empirical AI productivity literature finds consistent gains of 15–50% in task completion time across writing, software development, customer support, accounting, law, and translation. Benefits are disproportionately large for less-experienced workers, compressing skill gaps within occupations. Despite rapid adoption — nearly 40% of U.S. adults used AI tools by late 2024 — aggregate labor market indicators through 2024–25 show limited disruption, mostly concentrated in entry-level segments of highly exposed occupations.

Jan 22, 2026 · Alexander Kustov
Political scientist Alexander Kustov argues academia faces structural collapse from AI, not incremental change. AI can already produce publishable first-quartile social science research: Tibor Rutar generated a full paper from prompts alone, Paul Novosad replicated the feat in 2–3 hours. Kustov’s ten theses hold that the research production bottleneck is gone, the publishing system isn’t built for this volume, and most academics — the most dispositionally conservative institution on the planet — haven’t noticed the ground shifting.

Jan 15, 2026 · McKinsey
McKinsey announced a workforce of 20,000 AI agents and embedded an AI competency test into its graduate hiring process. The agents handle knowledge synthesis and research firm-wide. Requiring AI aptitude in new hires marks a shift in what the firm is selecting for: the ability to direct and work alongside agents is now part of the baseline expected of incoming consultants, alongside traditional analytical and interpersonal skills.

Jan 2, 2026 · Erik Brynjolfsson
Stanford Digital Economy Lab director Erik Brynjolfsson argues 2025 planted the seeds of a profound transformation, with AI agents now handling execution while humans focus on problem definition and evaluation. A PwC survey found 79% of companies leveraging agentic AI. Brynjolfsson frames the shift through his research on tacit knowledge: AI has learned to execute complex action sequences without explicit rules, finally automating what was considered irreducibly human.

Dec 3, 2025 · Anthropic
Anthropic announced that Claude Code crossed $1 billion in revenue and simultaneously acquired Bun, the fast JavaScript runtime built by Jarred Sumner. Bun’s toolchain joins the Claude Code ecosystem, accelerating infrastructure for agentic development workflows. The $1 billion milestone arrived in roughly two years. That is a measure of how fast enterprise developers moved from evaluating AI coding tools to depending on them for core engineering work.

Feb 4, 2025 · (Multiple)
Most for Junior Workers* — A randomized study of 5,172 customer support agents published in the Quarterly Journal of Economics found AI assistance raised productivity by 15% on average, measured by issues resolved per hour. Gains were largest for lower-skilled and newer workers, who improved both speed and quality; top performers saw small speed gains with slight quality declines. AI also facilitated worker learning, improved English fluency among international agents, and made customers measurably more polite.

Jul 13, 2023 · (Multiple)
Most for Weaker Workers* — A preregistered experiment recruited college-educated professionals to complete incentivized writing tasks with and without ChatGPT. Workers assigned AI finished faster, produced higher-quality output, and reported enjoying the work more. Workers with weaker baseline skills benefited most. Published in Science, the Noy and Zhang study was among the first rigorous demonstrations that generative AI could raise skilled professional productivity — and that gains would flow unequally, compressing skill differences within occupations.

Read the full log:

The week ending March 15, 2026 sharpened a question the previous week raised: what is AI actually doing to work, versus what companies are saying it is doing. Amazon called an emergency engineering meeting after four severe site outages, with an internal document citing “Gen-AI assisted changes” as a trend of incidents since Q3, then quietly deleted the reference before the meeting began. Oracle raised its restructuring fund to $2.1 billion and announced further cuts, but Fast Company’s analysis suggests the bigger driver is the company’s $50 billion data center buildout rather than AI replacing workers directly.

The organizational imagination may be running ahead of the disruption. Box CEO Aaron Levie argues enterprises will eventually field 100 times more agents than people, with every contract review, financial audit, and customer support case flowing through agents. Notion co-founder Simon Last describes engineers already making the shift from individual code authors to managers of multi-agent systems that generate, review, and fix each other’s output. A Udacity survey found only 9% of executives actually want to replace their entire workforce with AI; human creativity, customer preference, and legal liability are the limits AI hasn’t cleared.

Mar 12, 2026 · Oracle
Oracle raised its restructuring fund to $2.1 billion and is cutting thousands of jobs, but Fast Company’s analysis found the bigger driver is the company’s $50 billion AI data center buildout rather than AI directly replacing workers.

Mar 12, 2026 · Notion
Notion co-founder Simon Last describes how engineering teams are reorganizing around AI agents, with developers moving from individual code authors to managers of multi-agent systems that generate, review, and fix each other’s output with limited human intervention.

Mar 10, 2026 · Amazon
Amazon called an emergency engineering meeting after four severe site outages in a week. An internal document cited “Gen-AI assisted changes” as a “trend of incidents” since Q3, though the reference was quietly deleted before the meeting began. Amazon said the incidents had a “high blast radius.”

Mar 8, 2026 · Aaron Levie (Box)
Box CEO Aaron Levie argues that agents will become the primary users of software, with enterprises eventually fielding 100 times more agents than people. Every contract review, financial audit, customer support case, and line of code will flow through agents, making software designed for humans effectively obsolete.

Mar 8, 2026 · (Multiple)
As companies like Snowflake deploy AI agents that handle monitoring, on-call response, and task assignment, hierarchies are reshaping: agents now assign work to humans and, in some firms, inform decisions about promotions and layoffs. White-collar jobs, especially entry-level roles, are disappearing fastest.

Mar 6, 2026 · (Multiple)
A Udacity survey found only 9% of executives and managers want to replace their entire workforce with AI. Most cite human creativity, customer preference for human interaction, and legal liability as limits AI can’t clear. Companies rushing to cut headcount in favor of agents are likely to meet resistance across the org chart.

A couple of things I found this week that backfill gaps in the full log

Feb 19, 2026 · Mike Konczal
An economist describes integrating terminal AI tools like Claude Code into real-time economic data analysis: what used to require an hour of prep before 8:30 a.m. data releases now takes seconds, with robustness checks and chart generation handled autonomously. The judgment about what to measure remains human.

Aug 18, 2025 · MIT Sloan Management Review
AI coding tools can make developers up to 55% more productive, but MIT researchers warn that rapid deployment creates dangerous technical debt, especially in legacy systems. AI-generated code can destabilize architecture in ways that only surface months later, potentially erasing productivity gains and crippling scalability.

Read the full log:

Writing and speaking are acts of influence — your attempt to change other people’s behavior. I take whatever you say and write as your attempt to show up in the world and change things. I judge you for what that entails, not so much for your words and punctuation.

I don’t take what you say and write as a pure expression of your true self, whatever that means. So, I am not offended when your writing and speech carry patterns picked up from elsewhere.

You picked up writing and speech patterns from your mom, the biased one. And from your dad, the racist one. And from your 5th-grade teacher, the one who belittles her daughter. And from your pastor, the one who underpays his secretary. And from the trainer at the gym, the one who can’t see past your skin. And from your lack of protein at breakfast. And from the sinus infection that kept you awake all night. And the amount of lead paint in the window trim at your grandfather’s house. And a school district in 1980 that refused to fund the arts program at your elementary school. You and your actions are the nexus of a million imperfect lives and circumstances.

But, Jeremy, wouldn’t you care if a journalist used AI to report the news? I would care if they misreported the facts, whether or not they used AI.

But, Jeremy, wouldn’t you care if a job applicant used AI to write code they couldn’t write themselves? I would care if they couldn’t reliably reproduce it, the same way I would care if a teammate’s health declined and they could no longer perform their job.

But, Jeremy, don’t you care that the AI writer is lazy? Sometimes. I also care that you, too, are lazy sometimes, but I don’t dismiss you for it.

But Jeremy, wouldn't you care if a priest wrote with AI rather than divine inspiration? I would care if the priest's words led their congregation toward cruelty. That's what divine inspiration is supposed to prevent, and it's the only measure I have access to.

I care about what it is you want to do in the world and how you want to change it. I infer those from what you write and say, and from other behaviors. If AI or bad goat cheese gave you some of the words, I don’t care.

If AI’s transformation of work interests you, you can read about that here:

Welcome to the Treeline! The view opens up here. This update adds new entries to my log of AI’s transformation of work, the good and the bad.

New entries

Mar 7, 2026 · Benzinga
Salesforce’s CEO dismissed AI layoff fears on Wednesday; by Friday the economy had shed 92,000 jobs and Morgan Stanley, Oracle, and Capital One had all announced cuts, complicating his argument.

Mar 5, 2026 · Oracle
Oracle is planning thousands of layoffs across multiple divisions, some targeting categories expected to shrink due to AI, as the company faces a cash crunch from its massive AI data center expansion.

Feb 19, 2026 · IBM
While other companies cut headcount in favor of AI, IBM is tripling entry-level hiring across engineering, HR, and support, arguing that companies who hollow out junior pipelines now will struggle in five years.

Subscribe now

Read the full log at

Welcome to the Treeline! The view opens up here. This update adds new entries to my log of AI’s transformation of work, the good and the bad.

There is so much to keep up with! This update adds 12 new entries to the log, covering Scientific American, Anthropic (Wiz customer story), Anthropic (Rakuten customer story), and 9 more.

Mar 3, 2026 · Scientific American
Studies find AI-assisted developers ship more code but log longer hours and face more rollbacks, complicating the productivity narrative.

Mar 2, 2026 · Anthropic (Wiz customer story)
Wiz used Claude Code to complete a language migration estimated at two to three months of specialized engineering in roughly 20 hours, with 90% of engineers now using it daily.

Mar 2, 2026 · Anthropic (Rakuten customer story)
Rakuten achieved seven hours of sustained autonomous coding and a 79% reduction in time to market using Claude Code across its engineering organization.

Mar 2, 2026 · Anthropic (Ramp customer story)
Ramp deployed Claude Code across engineering, with 50% weekly active usage and up to 80% faster incident investigation through AI-built internal tooling.

Mar 2, 2026 · Anthropic (n8n customer story)
n8n built its AI Workflow Builder on Claude, cutting 80% of the manual work in assembling workflow automations and letting users go from idea to working workflow in minutes.

Mar 1, 2026 · OpenClaw
Anthropic’s open-source agent framework went from zero to 250K+ GitHub stars in under four months, overtaking React and Linux on the all-time leaderboard.

Mar 1, 2026 · Lenny’s Podcast
Anthropic’s head of design Jenny Wen says the classic design process is “basically dead” as engineers with multiple Claude agents ship features faster than traditional research-and-mockup cycles allow.

Feb 19, 2026 · Business Insider
Salesforce’s internal survey shows most employees feel AI increases their productivity, though fewer report reduced workloads, as the company bets its future on AI agents.

Feb 15, 2026 · Business Insider
A six-person startup produced output that would have required 20 to 30 engineers five years ago; Canva’s CTO calls the agent-driven shift “really impressive.”

Jan 29, 2026 · Alex Imas
A living review of micro and macro evidence finds AI productivity gains, long documented in individual studies, are now appearing in aggregate economic data for the first time.

Oct 30, 2025 · Brex
A content designer with no coding background now builds Figma plugins and ships features using Claude Code, illustrating how AI extends engineering capability to non-engineers.

Apr 1, 2025 · Stanford HAI
Stanford’s annual AI report finds inference costs dropped 280-fold since 2022, hardware costs fell 30% annually, and open-weight models closed the gap with closed models to just 1.7%.

You can access the entire log, too.

Welcome to the Treeline! The view opens up here. This update adds new entries to my log of AI’s transformation of work, the good and the bad.

This update adds 4 backfill entries to the log, covering PwC, Goldman Sachs, Klarna, and 1 more between 2023-2025.

New Entries

Dec 19, 2025 · PwC
PwC uses Harvey AI to automate legal entity reviews and liquidation workflows, shifting M&A support from manual checklists to end-to-end AI-driven execution across global projects.

Aug 13, 2025 · Goldman Sachs
Over 50% of Goldman’s 46,000 employees adopted the internal GS AI Platform, yielding 20% productivity gains for coders and a 15% drop in post-release bugs, with CEO targeting 100% adoption by 2026.

Feb 27, 2024 · Klarna
Klarna’s OpenAI-powered assistant handled 2.3 million conversations in its first month, doing the work of 700 full-time agents with equal satisfaction scores and an estimated $40 million profit improvement.

Aug 2, 2023 · McKinsey
McKinsey’s generative AI platform Lilli reached 72% firm-wide adoption, with over 500,000 prompts monthly and colleagues reporting up to 30% time savings in searching and synthesizing knowledge.

Read the full living brief: