For two years the story has been the same. AI writes code faster. Agents close tickets. One engineer does the work of three. Pick your multiplier.

If you've read Mats's post on running four Claude instances across separate worktrees, you've seen the setup that got me here. This is the part that surprised me after a few weeks of living in it.

When I started running multiple Claude Code agents in parallel, each in its own git worktree, the code stopped being the constraint almost overnight. I could have four branches moving at once. What I could not do was review four branches at once, hold the architecture of all of them in my head, or trust that work I didn't read still made sense. The typing got cheap. Everything around the typing got expensive.

That's the part nobody warns you about. AI doesn't remove the bottleneck. It moves it somewhere you weren't looking.

Where the work actually went

Generating code was never the slow part of building software. The slow parts were deciding what to build, agreeing on how, and being confident it was right. We spent years pretending those were soft skills sitting next to the real engineering. They were the engineering. The code was the easy 20% that felt like 80% because it was the only part you could see yourself doing.

Take that part away and what's left is the work that was always hard: judgment, taste, knowing when "it runs" and "it's correct" are different sentences. An agent will happily produce a function that passes the test you wrote and quietly violates an assumption you never wrote down. The faster it generates, the faster those gaps pile up.

Reading is the new writing

I now spend more time reading code than writing it, and the gap keeps widening. Reading code you didn't write, that arrived faster than you can think about it, is a genuinely different skill than authoring it yourself. It's closer to reviewing a pull request from a brilliant colleague who is also occasionally, confidently wrong and never tired.

Most engineers are not trained for that volume. We optimized careers around producing. The next few years reward verifying. The engineer who can look at a large diff and instantly find the one wrong assumption is worth more than the one who can produce that diff a little faster than the machine.

What this means for teams in 2026

If you hand a team AI agents and change nothing else, you don't get a faster team. You get the same team drowning in plausible-looking output nobody has the capacity to vet. Throughput at the keyboard goes up. Throughput of trustworthy, shipped, understood software does not move, and sometimes it gets worse.

The teams pulling ahead aren't the ones with the best prompts. They're the ones who rebuilt their process around the new bottleneck. Smaller, more frequent changes so review stays possible. Tests and types as the contract the agent has to satisfy, not an afterthought. Architecture written down so the assumptions live outside someone's head. The cheapest assumption to catch is the one you wrote down before the agent started, not the one you spot in a diff afterwards. A culture where "I didn't read it closely, but it works" is not an acceptable sentence.

None of that is new advice. That's the point. AI didn't invent new engineering discipline. It removed the slack that let us get away without it.

The skill that compounds

If I were starting today, I would not spend my energy getting faster at producing code. The machine wins that race and the gap only grows. I would get very good at the thing the machine can't do: holding the whole system in my head, knowing what correct looks like, and spotting the quiet wrong assumption before it ships.

And to be clear, reading faster is the coping skill, not the cure. The teams that really escape the race against the clock are the ones doing the work upstream, capturing intent and treating the spec as the contract, so there's less to catch in the first place. That's a post of its own.

The bottleneck moved. The engineers who move with it will look, a year from now, like they have some unfair advantage. They won't. They'll just have noticed earlier where the real work went.