Gergely Orosz: What AI Changed and Broke in Software Engineering in 2026

Gergely Orosz of The Pragmatic Engineer reports on what AI changed, left alone and broke in software engineering over the past twelve months, based on visits inside OpenAI, Anthropic, Cursor and Ramp. In The state of the tech industry in 2026, a 29-minute keynote at LDX3 New York, Orosz moves through sixteen changes, five things that held, a list of what is failing, and advice for engineering leaders who are wondering whether to stay.

TLDW

At the companies Orosz visited, almost no code is written by hand, engineers run five to ten agents at once, the IDE is fading, and migrations that took years now take months. Teams, planning for complex work, and testing have not changed, and non-engineers are still not shipping production code. What broke is code review, software quality, hardware supply, engineers’ focus, and the job of engineering leadership, which many are leaving. Orosz expects cloud agents, shipping code nobody reads, and hiring that screens for AI fluency, and tells leaders to get hands-on again.

Thoughts

Code review is dead and nothing has replaced it. That is the uncomfortable finding in this talk. Orosz calls what is left “zombie code reviews”: the volume of agent-written changes is so high that people approve without reading, and one startup is phasing reviews out because the ritual no longer does anything. Ramp has told engineers it will trust them to skip review on non-critical code. In the next breath Orosz says quality is visibly worse across software in everyday use. Those two facts belong together. The industry removed a control because it could not keep up, and has not yet agreed on what stands in its place. The teams doing this well, on the evidence here and elsewhere, moved the check earlier (tests, lint rules, agents that verify their own work) and later (sampling what shipped). Most teams have simply stopped looking.

The list of what did not change is the best corrective to the hype I have seen from someone with this much access. Anthropic still runs two-pizza teams that own software and carry the pager. Planning for complex infrastructure looks as it did years ago, because building the wrong thing is still a waste even when building is fast. The Bun team still spends about as much effort on tests as on code, with AI writing both. And after asking around for a week, Orosz could not find one company where non-engineers push production code. They file bugs in Slack, and sometimes a bot opens a pull request. The story that everyone becomes a developer is, for now, a story.

The human cost gets two honest minutes. An engineer at Linear told Orosz that when AI arrived, doing in an hour what took colleagues a day felt great. Now that speed is the baseline, there is always another agent waiting for an answer, and the job feels like more work than before. Productivity gains that become expectations stop feeling like gains. The same pressure reaches leaders. Orosz spoke to 20 to 25 who are taking or planning a career break, citing founders who want half the team cut, or revenue doubled because AI spend doubled. Many now work as fractional CTOs and do not want to go back full-time. An industry this pleased with its tools is losing a notable number of the people who run it.

Some care is needed with the sample. The sixteen changes come from OpenAI, Anthropic, Cursor, Ramp, Linear and similar companies, which are the most AI-forward engineering teams in the world, and Orosz says as much about Anthropic. “Close to 100% AI-generated code” is true of startups that track the number. It is not a census. Still, a few of the signals come from hard data and not from anecdotes: Linear now sees more tickets created by agents than by people, GitHub’s agent-only pull requests rose nearly tenfold between January and August, and Cursor’s own team calls its IDE a legacy product. When the company that won the editor war says the editor is shrinking, that is worth more than a survey.

The most useful idea for anyone outside engineering arrives near the end, borrowed from Titus Winters, an author of Software Engineering at Google. Success takes intelligence (knowing how to do it), wisdom (knowing what to do) and charisma (convincing others to do it). AI is making the first one cheap, so the other two matter more. That fits everything else in the talk. Deep domain knowledge is in demand, specializations in particular languages and platforms are disappearing, and the advice to leaders is to build things again and not to outsource their own learning. It sits awkwardly beside the prediction that we will soon ship code nobody reads. Wisdom about what to build is hard to keep if you stop looking at what was built.

Key Takeaways

The turning point was January, when engineers came back from the winter break to find the newest models and their harnesses were good enough to write nearly all the code.
Running five to ten agents in parallel is now common among very experienced engineers, including at Anthropic, Linear and Cockroach Labs.
The heavyweight IDE is fading. Cursor moved to an agent interface in April, JetBrains is pivoting to a harness called Air, and OpenAI decided against forking VS Code for Codex.
Nearly every serious tech company has built its own coding harness, often to connect internal data it will not hand to an outside tool. Orosz names Ramp, Stripe, Uber, Block, Shopify, Google, Meta and Amazon among others.
Development work often starts in Slack, by tagging an agent and asking it to implement something.
Migrations are fast. Orosz says OpenAI is about 90% through moving its API from Python to Rust after four to five months, Airbnb replaced a UI testing library in six weeks, and Uber moved from JUnit 4 to 5 in four months.
Cost became a crisis in the spring, when Uber was reported to have used up its annual AI budget. By September, large companies had cut per-token costs by about half with open models and smarter routing.
Specializations are dissolving. OpenAI stopped hiring specifically for iOS and Android engineers this year, and projects are increasingly done by one or two engineers because each person’s agents already collide.
Teams are smaller and junior hiring keeps falling slowly.
Old engineering vocabulary makes agents better. Matt Pocock found that terms from The Pragmatic Programmer and A Philosophy of Software Design, such as tracer bullets and deep modules, improve the code an agent writes.
There is a CPU shortage, not only a GPU and memory shortage. Lead times for servers at cloud providers have gone from one or two weeks to as long as six months, and some regions are taking no new orders.
AI startups pay far more than others, and Orosz says Anthropic is hiring CTOs of major public companies as individual contributors.
Companies that are maturing fastest start from a business outcome and build agent systems at team or company level. Automating individual tasks is the easy and less useful path.

Chapters

00:00 Inside the Labs, Big Tech and Startups

Orosz sets out the sources: recent visits to the headquarters of OpenAI, Anthropic, Cursor and Ramp, and long conversations with people at Uber, Linear and other startups. The talk has four parts: what changed, what did not, what broke, and what comes next.

01:01 Nobody Writes Code by Hand, and Everyone Runs Parallel Agents

Engineers still edit by hand, but generation is nearly all AI at startups that measure it. Boris Cherny described five local Claude Code sessions plus five to ten on the web. An engineer at Linear put it as the end of one mouse, one keyboard, one screen. Peter Mattis of Cockroach Labs named five to ten sessions as a personal cognitive limit.

03:02 The IDE Fades and Every Company Builds a Harness

Orosz did not expect this one. The last major VS Code fork shipped in November, and the tools that followed look like agent consoles. Kent Beck’s comment was that people still need context for their decisions, but the context has changed. Meanwhile companies of every size are building internal harnesses, some on top of Codex, Claude Code or OpenCode.

05:04 Slack, the Agentic Software Factory and Fast Migrations

Orosz once treated “agentic software factory” as a buzzword and now describes OpenAI’s, where agents monitor production and raise pull requests to improve it, with humans reviewing the critical ones. It is not fully automated and is not meant to be. The migration examples follow, including one engineer at Anthropic moving a codebase from Zig to Rust.

08:09 Agent-Made Tickets, Costs and Vanishing Specializations

Linear’s data shows agent-created issues passing human-created ones for the first time. GitHub reported 7.7 million agent-generated pull requests in January and close to ten times that by August. Skills usage at Factory AI rose from about 30% of users in February to over 80%. The section ends with costs, the end of platform specialists, and smaller teams.

12:12 What Did Not Change

Orosz stresses that all sixteen changes happened in twelve months, then lists what held. The team remains the unit of work, because people go on vacation, inspire each other and keep each other accountable. Planning remains for complex work. Testing keeps its share of effort. Non-engineers do not ship to production. And old design ideas are being rediscovered because they help agents.

16:13 What Broke: Reviews, Quality, Hardware and Focus

Pull requests, commits and new repositories on GitHub are up about five times over three years, with total pull requests doubling in the last two months alone. Reviews cannot absorb that. Orosz sees paper cuts across products from Spotify to Substack. Companies are paying now for CPU capacity that will not arrive until December. Engineers report constant context switching.

20:16 Why Engineering Leaders Are Walking Out

The reasons Orosz gathered: unrealistic demands from founders, startups that are losing and equity going to zero, companies too slow to adopt AI (which leaves a leader’s skills stale), fewer management roles as teams shrink, far better pay at AI companies, a good moment to start something, and plain burnout.

22:18 What Comes Next

Cloud agents and harnesses come first. Ramp built its own for three reasons: local machines are limited, front-end tooling is better in the cloud, and it wanted remote development environments anyway. Next, Orosz expects teams to stop reading code, citing a conversation with Charity Majors. Evals join CI/CD, a refactoring wave arrives, and hiring screens for AI fluency and a positive attitude toward it.

27:22 How to Thrive as an Engineering Leader

Get hands-on and stay there. Will Larson, running a 50-person engineering organization, has shipped more code in 15 months than in the previous five to ten years. Build the AI infrastructure your team needs. Use AI to remove friction, and do not outsource your learning. Accept less people management for now. Then ask why you got into tech in the first place.

Notable Quotes

“The IDE is now just a legacy product. They need it for enterprise, but it’s not growing and it’s shrinking.”
Gergely Orosz, relaying what the Cursor team said about its own editor

“Building the wrong thing is still a waste even when building is faster.”
Gergely Orosz, on why planning survives for complex work

“If your team is still doing code reviews, you’re probably not doing code reviews.”
Gergely Orosz, on what the talk calls zombie code reviews

“The job of an engineering leader just got a lot worse, a lot harder.”
Gergely Orosz, on why leaders are taking career breaks

“It’s not an if, it’s a when. It’s coming.”
Gergely Orosz, on shipping code without reading it

“I feel I’ve learned more in the past year than in the previous 5 years combined.”
Peter Mattis of Cockroach Labs, as quoted by Orosz, on going back to hands-on work

The talk is dense and moves quickly, and the slides carry charts this summary cannot. Watch the full keynote here.