Hardly anyone writes code by hand anymore, yet two-pizza teams remain: Gergely Orosz on a year of change in software engineering
Nov 13, 2026Original LeadDev LDX3 뉴욕(게르게이 오로스) · The Pragmatic Engineer · McKinsey ↗
In 3 lines
- AI writes nearly all the code, and engineers run 5 to 10 agents in parallel
- Code reviews have become theater, and quality and focus are suffering
- Teams, planning and tests remain, and no company had non-engineers shipping to prod
Gergely Orosz, who writes The Pragmatic Engineer newsletter for over 1.1 million subscribers, gave the keynote at LDX3 New York this fall to 2,000+ engineering leaders. After visiting OpenAI, Anthropic, Cursor and Ramp and getting unpublished data from GitHub and Linear, he sorted the last 12 months into what changed, what broke and what stayed the same. We read the keynote, which LeadDev posted on YouTube on Nov. 10, alongside his written version and a McKinsey talk on how large companies are doing.
Why did a booming Meta shake up its engineering org?
Orosz opened with Meta. According to Reuters, Zuckerberg and his top lieutenants ran scenarios at a January leadership retreat in Hawaii that explored cutting team sizes by as much as 60%, because AI-native companies were doing with three to five people and agents what Meta did with 10 to 20 engineers, PMs and data scientists. From people inside, he heard that 10% of staff were let go and 20 to 30% were reassigned to data labeling overnight. Others started interviewing elsewhere and turned down on-the-spot retention bonuses of up to $1 million.
Hand-written code is rare; 5 to 10 agents is normal
He dates the shift to January, when Opus 4.5 and GPT-5.2 arrived with good harnesses; at most startups that track it, AI-generated code is now close to 100%. Claude Code creator Boris Cherny runs five terminal tabs plus 5 to 10 Claudes on the web, and Cockroach Labs cofounder Peter Mattis said "my cognitive overhead is about 5-10 agent sessions concurrently." Linear's Dima Zaytsev put it this way: "It used to be a simple world: one mouse, one keyboard, one screen," and now he rotates between 5 to 10 worktrees.
Boris Cherny, Peter Mattis, Dima Zaytsev · The Pragmatic Engineer ↗Agent-authored PRs overtook humans
Data GitHub shared with Orosz shows agent-authored pull requests grew ninefold in eight months, and in August 2026 there were more agent-authored PRs than human-authored ones. On Linear, agent-created issues have outnumbered human-created ones since July. At Factory AI, 83% of users use agent skills, up 2.5x from February. Orosz estimates there are now likely 75M+ fully AI-generated PRs a month, about 3x the 25M human PRs of December 2023.
Gergely Orosz · The Pragmatic Engineer ↗The IDE fades, and every company builds its own harness
Cursor relaunched in April without the IDE interface, and its team told Orosz the VS Code fork is now a legacy product. Codex launched as a non-IDE in February, and JetBrains is pivoting to Air, an agentic environment. Kent Beck told him: "It's not that we don't need more perspective and context for our human-based decisions; it's that the context has changed." Most mid-sized-and-above companies have built their own agent harness (Stripe's Minions, Block's Goose, Shopify's River, Ramp's Inspect and more), and at OpenAI work often starts in Slack with "@Codex, implement this."
Gergely Orosz · The Pragmatic Engineer ↗Migrations that took years now take weeks
Anthropic migrated the package manager Bun from Zig to Rust in 11 days, versus an estimated 1.5 engineering years. OpenAI is moving its API layer from Python to Rust and is about 90% done after roughly five months. Airbnb migrated 3,500 end-to-end test files in six weeks, and Uber moved 600,000 JUnit 4 tests across 15 million lines of code to JUnit 5 in four months. Orosz expects a golden age of refactoring and rewrites that had been put off for years.
Gergely Orosz · The Pragmatic Engineer ↗One or two engineers per project, and blurring specializations
Katelyn Lesse, Head of Engineering for Claude Platform at Anthropic, said an individual project often can't have more than two people, because each engineer already runs several agents that are "stepping on each other's toes." ChatGPT's head of engineering said that last year they were still hiring iOS and Android specialists and this year they are not. Orosz sees falling demand for mobile and frontend, rising demand for AI engineers and FDEs, and junior hiring that keeps sliding without collapsing. He stressed that apart from junior hiring, none of these 16 changes were true a year ago.
What broke: code review became theater
An engineer at a mid-sized startup told him: "Everyone is playing the theater of doing reviews." With the volume of changes, people take the path of least resistance and stamp LGTM on everything, a line that drew the strongest reaction at LDX3. Ramp now trusts engineers to skip reviews on non-critical code. Product quality feels worse across the board, and ordering cloud CPUs has gone from one to two weeks to as long as six months. Dima Zaytsev said that once he did in an hour what took peers a day, but that is now the baseline expectation, so he ends up with more work than before AI.
What stayed: teams, planning, tests
Katelyn Lesse pushed back on the idea of "two humans and a bunch of agents": she still has teams that own software and on-call, and they are still two-pizza teams. Complex infrastructure work still gets long planning. Bun creator Jarred Sumner said AI writes nearly all code and tests, yet testing still takes about as much time as writing code, just like before AI. Orosz asked AI labs and startups and could not find a single company where PMs, designers or non-technical people ship to production. Matt Pocock found that telling agents to use "tracer bullets" from The Pragmatic Programmer produced better code.
Katelyn Lesse, Jarred Sumner, Gergely Orosz · The Pragmatic Engineer ↗Leaders step away, and builders get hands-on again
Orosz spoke with 20 to 25 engineering leaders who were taking, or about to take, a career break, citing unrealistic founder expectations, fear of being stuck in slow-moving organizations, and the urge to build again. His advice to leaders is to stay hands-on: Will Larson, who leads a 50-person engineering org, shipped more code to production in the last 15 months than in the previous five to ten years. Titus Winters, lead author of Software Engineering at Google, told him that intelligence is knowing how, wisdom is knowing what, and charisma is convincing others, and when intelligence becomes commonplace the other two matter more. Orosz added that wisdom is most easily earned through domain expertise.
Counterpoint: most large companies are not at 2x yet
Around the same time, McKinsey shared a survey of a few hundred large enterprises. Most still report modest gains, and only the top 20 to 25% see roughly 2x end-to-end improvements. Those leaders replace two-week sprints with one- to two-day agent-driven cycles and split teams into pods of two to four, which McKinsey says does not necessarily mean cutting headcount. Readers in the comments on Orosz's post asked whether the companies he talked to have incentives to overstate AI's impact, and whether more PRs mean more revenue or profit. For tired leaders, Orosz passed on Peter Mattis's advice: "Remind yourself why you got into this in the first place. I got into software engineering because I like building stuff. Now I can build faster."
McKinsey, Pragmatic Engineer readers · YouTube · AI Engineer ↗
