The Code Delusion
AI coding assistants made developers feel faster without improving delivery. Agents are different, and they shift the value of engineering to design.
Two years of feeling faster
From 2023 through 2025, the software industry ran an experiment on itself. Satya Nadella estimated that 20 to 30 percent of the code in Microsoft's repositories was written by AI. Sundar Pichai said more than a quarter of Google's new code was. GitHub Copilot passed 20 million all-time users, and Microsoft says 90 percent of the Fortune 100 use it. The industry declared victory before anyone checked whether the software actually got better.
The evidence says it mostly did not. Google's 2025 DORA report found that 90 percent of software professionals now use AI at work. Faros AI, in its AI Productivity Paradox report, studied telemetry from more than 10,000 developers and found that on teams with heavy AI use, individual output was up: 21 percent more tasks completed and 98 percent more pull requests merged. But those pull requests were 154 percent larger, review time grew 91 percent, bugs per developer rose 9 percent, and Faros saw no significant link between AI adoption and improvement at the company level. A CodeRabbit analysis of 470 open-source pull requests found about 1.7 times as many issues in AI co-authored changes as in human ones, with cross-site scripting flaws 2.74 times more common.
The study I find most telling came from METR, which ran a randomized controlled trial with 16 experienced open-source developers across 246 real tasks in their own repositories, mostly using Cursor with Claude 3.5 and 3.7 Sonnet. Beforehand, they expected AI to make them 24 percent faster. Afterward, they believed it had made them 20 percent faster. Measured, they were 19 percent slower. It is a small study, it has not been peer reviewed, and its authors caution against generalizing from it. Still, the gap between what the developers felt and what the clock recorded is the part I keep coming back to. A lot of the industry's confidence in that period rested on how the tools felt.
Then the tools changed. In January 2026, Anthropic released Cowork, a desktop agent for people who are not engineers, as a research preview. At the end of that month it added plugins, including one for legal work. Over two trading days in the first week of February, roughly $285 billion in market value came off software, financial services and asset management stocks. LegalZoom fell almost 20 percent, and Thomson Reuters, RELX and Wolters Kluwer dropped sharply. That same week, Claude Opus 4.6 shipped with a one-million-token context window in beta and teams of agents in Claude Code. Markets overreact, and there was a partial rebound. But investors had read the moment the way I did: the tools arriving now are a different kind of technology from the autocomplete assistants that came before. The assistants helped developers type faster. These tools aim to do the typing themselves.
Why I expect this to continue
Most arguments about AI coding are about whether today's tools are good enough. I think the more useful question is the trajectory. For several years, model performance has improved fairly predictably as labs add compute, data and parameters, and I am not aware of convincing evidence that this has stopped. There is now a second lever as well: models that are given more time to reason through a hard problem do better on it. Neither trend is a law of nature, and either could slow. But planning as if both will stop soon seems like the riskier bet.
Context windows are part of this. At 200,000 tokens, a model holds a few files. At one million, it can hold a sizable codebase in a single session: the dependencies, the edge cases, the design decision buried in a file nobody has touched in two years. If windows keep growing, a model will eventually be able to read a full monorepo along with its history, its design documents and its issue tracker. At that point it can see what the code does and also much of the reasoning behind why it was written that way.
Cost is moving the same way. Epoch AI found that the price of reaching a given level of model performance has been falling by between 9 and 900 times per year, depending on the task. Very few technologies have ever become cheaper that quickly. Devin, the autonomous coding agent, cut its entry price from about $500 a month to $20 with its 2.0 release. Each cut of that size puts capable coding agents within reach of teams that could not have justified them before.
The development I find most consequential is that these tools are now used to build themselves. Boris Cherny, who created Claude Code at Anthropic, has said he has not written a line of code by hand since November 2025, and that Cowork was built in about a week and a half, with Claude Code writing nearly all of it. When better tools shorten the time it takes to build the next version of the tools, each cycle can arrive sooner than the last. How far that loop runs is an open question, but it has clearly started.
METR, the same group that ran the slowdown study, also tracks a different measure: the length of software tasks, counted in human working time, that an AI system can complete with 50 percent reliability. From 2019 to 2025 that horizon doubled roughly every seven months, and METR notes the pace may have picked up in 2024. Its estimate for Claude Opus 4.5 was about four hours and 49 minutes, with a wide margin of error. If the trend holds, day-long tasks come next, and then week-long ones. This is observational data from an organization with no product to sell and a record of publishing results the industry did not enjoy. I take it seriously for that reason.
What is already happening
On February 26, 2026, Jack Dorsey announced that Block would go from more than 10,000 employees to just under 6,000. His shareholder letter was direct: "Intelligence tools have changed what it means to build and run a company." He said the cuts were not a response to financial trouble, and that he expected most companies to reach the same conclusion within the next year. Block's shares rose more than 20 percent in after-hours trading. More than 4,000 people lost their jobs in that decision, and I do not want to write past that. But it matters that the market rewarded a 40 percent workforce reduction at a company that said it was doing well.
Other leaders have been saying similar things. Tobi Lütke told Shopify staff in an internal memo: "Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI." Marc Benioff said Salesforce would not hire new engineers in 2025, citing a 30 percent productivity gain in engineering, and told a Davos panel: "We are the last CEOs who are only going to be managing humans as our workforce." Dario Amodei of Anthropic said at Davos in January 2026 that we might be six to twelve months away from models doing most, maybe all, of what software engineers do end to end. Mark Zuckerberg told Dwarkesh Patel he expects most of the code for Meta's AI efforts to be written by AI within 12 to 18 months. These are people with something to sell, and some of these forecasts will prove early. I still find it notable how consistently they point the same way.
Andrej Karpathy, who coined the term "vibe coding," has already moved past it. He described giving an agent a single prompt with access to his DGX Spark and a setup job that, by his account, would have taken him a weekend a few months earlier. The agent ran for about thirty minutes. It hit problems, searched for solutions, debugged its own work, and came back with everything running. He did not touch anything. His view is that the old picture of programming, a person typing code into an editor, no longer describes the work. He calls the new practice "agentic engineering": directing long-running AI systems that have tools, memory and instructions, often several at once. The job moves from using AI to write code toward managing AI that writes code.
Solo builders show the same shift at a smaller scale. Pieter Levels, an independent developer, built a multiplayer flight simulator with AI tools without having made a game before, and by his own account it reached $1 million in annualized revenue in 17 days. A viral launch is not a business model, and I would not treat it as one. I read it as a sign of how little now stands between an idea and a working product.
The labor data is starting to show it too. Stanford researchers working with ADP payroll records found that employment for software developers aged 22 to 25 was nearly 20 percent below its late-2022 peak by July 2025, while employment for more experienced workers in the same occupations held steady or kept growing. SignalFire reports that new graduates made up just 7 percent of Big Tech hires in 2024, with new-graduate hiring down more than 50 percent from 2019. The age groups moved together for years and then split after generative AI tools arrived. My reading is that the most exposed work is the part that consists of translating a clear specification into code, which is the work juniors were traditionally given.
Meanwhile, the demonstrations of long-running agents keep getting harder to dismiss. An Anthropic researcher had sixteen Claude agents build a C compiler in parallel, one capable of compiling the Linux kernel. Rakuten reports that Claude Code worked autonomously for seven hours on a change to a large open-source library and matched the reference implementation with 99.9 percent numerical accuracy. And a junior engineer at a large San Francisco tech company told the San Francisco Standard that all of his code is now written by AI:
"I'm basically a proxy to Claude Code. My manager tells me what to do, and I tell Claude to do it."
He described his main feeling as grief, at watching a skill he spent years building become available to anyone. I think that deserves more respect than it usually gets in these discussions.
Where I land
At BearPlex, what I look for in engineers is how they think about systems: how they break down an ambiguous problem, define constraints and failure modes, and design structures that keep working without someone watching. Typing speed is not what I look for. I do not want to overstate where we are. We are still learning how to work with agents.
I think a large share of what was called software engineering was closer to transcription: turning a ticket into a pull request, a design document into functions, someone else's architecture into code. That work was real and it took skill, but it is the work agents now do well, and I do not expect it to come back as a full-time job. Challenger, Gray & Christmas counted 54,836 announced US job cuts attributed to AI in 2025. Many of those cuts will turn out to be mistakes. Forrester reports that 55 percent of employers regret laying people off because of AI, and predicts that half of those layoffs will be quietly reversed, often offshore or at lower pay. My own guess, and it is only a guess, is that some of that regret comes from cutting by job title when judgment was what the company most needed to keep.
So I would ask a different question from "will AI replace software engineers?": how much of a company's engineering was design and judgment, and how much was implementation. The earlier assistants sped up typing and, at least in METR's sample, left experienced developers slower overall. The newer agents take over implementation and leave the design decisions with people, so they reward the engineers who were already doing that part of the job. If compute keeps scaling, if the tools keep improving the tools, and if prices keep falling, that gap widens. I cannot tell you where the ceiling is, and I have not seen anyone identify one convincingly.
A first-time game developer shipped something people pay for, a thirty-minute agent run replaced a weekend of setup, sixteen agents built a compiler, and a public company cut 40 percent of its staff and saw its shares rise. All of that has already happened. The companies under the most pressure are, I suspect, the ones that accumulated code without building systems around it. The durable asset in software is the architecture: the decisions about what gets built, how it gets verified, and why it exists at all. As the models improve, that kind of thinking gets harder to fake and more costly to lack.
Sources
- PCWorld: Microsoft CEO says up to 30% of its code is written by AIpcworld.com
- Entrepreneur: AI is taking over coding at Microsoft, Google and Metaentrepreneur.com
- TechCrunch: GitHub Copilot crosses 20 million all-time userstechcrunch.com
- Google: How are developers using AI? Inside the 2025 DORA reportblog.google
- Faros AI: The AI Productivity Paradox Report 2025faros.ai
- CodeRabbit: State of AI vs Human Code Generation reportcoderabbit.ai
- METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivityarxiv.org
- Axios: Anthropic's viral new work tool wrote itselfaxios.com
- The Batch: Claude Cowork plugins trigger a SaaS stock selloffdeeplearning.ai
- Reuters via Yahoo Finance: Anthropic's new AI tools deepen selloff in data analytics and software stocksfinance.yahoo.com
- Anthropic: Introducing Claude Opus 4.6anthropic.com
- Epoch AI: LLM inference prices have fallen rapidly but unequally across tasksepoch.ai
- VentureBeat: Devin 2.0 price cut to $20 per month from $500venturebeat.com
- Station F: Boris Cherny on not writing code by handstationf.co
- METR: Measuring AI Ability to Complete Long Tasksarxiv.org
- The Decoder: Claude Opus 4.5 and METR's time horizon estimatethe-decoder.com
- The Register: Block announces 40% layoffstheregister.com
- Yahoo Finance: Jack Dorsey lays off 4,000, says others will do the same within the next yearfinance.yahoo.com
- TechCrunch: Shopify CEO tells teams to consider using AI before growing headcounttechcrunch.com
- The Register: No new engineer hires this year, says Salesforcetheregister.com
- Axios: Salesforce chief predicts today's CEOs will be the last with all-human workforcesaxios.com
- Entrepreneur: Anthropic CEO on software engineering in 6 to 12 monthsentrepreneur.com
- Dwarkesh Patel: interview with Mark Zuckerbergdwarkesh.com
- The Decoder: Andrej Karpathy says programming is unrecognizable now that AI agents workthe-decoder.com
- Pieter Levels: fly.pieter.com reached $1M ARR in 17 dayslevels.io
- Stanford Digital Economy Lab: Canaries in the Coal Mine?digitaleconomy.stanford.edu
- SignalFire: State of Tech Talent Report 2025signalfire.com
- Anthropic: Building a C compiler with a team of parallel Claudesanthropic.com
- Rakuten case study: Claude Codeclaude.com
- The San Francisco Standard: AI writes the code now. What's left for software engineers?sfstandard.com
- Challenger, Gray & Christmas: 2025 year-end reportchallengergray.com
- The Register: Forrester on AI layoffs and rehiringtheregister.com

