Rejecting AI code even when it works
Vinicius Brasil on why he rejects AI-generated code he can't explain, even when it passes CI, and why review is now the real bottleneck.
Vinicius Brasil, describing his own work with coding agents, says big tasks still take him days, and that more often than not he rejects every change the agent made and starts over. The second run goes better than the first. He is clear the reason is not a newer model. It is that he has spent more time understanding the problem, which lets him drive the agent instead of being driven by it.
That inversion is the useful part of his argument. Implementation used to be the slow step. Now an agent can produce a large diff in minutes, so the slow step moves to review: not only your coworkers’ pull requests and whatever their agents wrote, but your own git diff after your agent finishes a job. Reviewing code you did not think through yourself carries a cognitive load that good process only partly reduces. Brasil follows the usual advice, plan mode first, big tasks split into phases, small changes shipped, and still finds reviewing unfamiliar generated code heavy.
He describes how he worked before agents. Given a task, he would read the codebase, consider several approaches, experiment, and implement only after that. Consolidating the context could take days, but by the time the pull request went up his confidence was higher and explaining each change to a colleague was easy.
He lists the conditions under which he rejects AI code even when it runs:
- when he cannot explain the approach in his own words
- when the diff is bigger than the problem
- when it introduces abstractions before they are proven necessary
- when it works locally but makes the system harder to reason about
- when he is trusting the output more than his own understanding
What ties these together is whether he understands the code, not whether it works. Code that runs and turns CI green can still be a bad solution, and Brasil’s point is that engineering has always been about adequate, scalable, extensible solutions rather than passing tests. He has seen engineers accept generated changes too quickly, which is why he argues for required human review alongside AI review, not in place of it.
None of this is an argument against the tools. He calls the agents impressive and uses them regularly. The claim is narrower: they still need a competent engineer steering them toward good solutions, and they cannot yet do that work autonomously in a way that holds up over time.
If you run teams that have adopted agents, the practical reading is about where you spend review attention. A green pipeline tells you the code runs; it tells you nothing about whether the solution is the right size for the problem or whether anyone on the team can explain it. Asking whether the author can describe the approach without reading the diff back to you is a cheaper and more honest gate than line coverage. The person who understood the problem before prompting will get more out of the same agent than the person who prompted first and tried to understand the output afterward.
Keep Reading
coding-agentsThe patch opens the attack window.
Armin Ronacher on the harness loop around coding agents: where it works, why he won't trust it with code he cares about, and what it means for security.
ai-codingspeed ships the flaw
Charity's essay on AI and engineering discipline: cheap regeneration turns code into a cache, so rigor has to move to production and validation.
rsyncmemcpy walks off the end of the receiver
A June 2026 statistical analysis of rsync releases finds the two carrying Claude-written commits are no buggier than the historical distribution.
Latest on the Wire
Full wire →- AI Coding Agents Boost Output But Code Review Bottlenecks IncreaseHacker News
- AI Overviews Cut Organic Traffic by 40%Hacker News
- Apache Patches Critical Roller RCE VulnerabilityBleepingComputer
- Complete Web Browser Built from Scratch in PythonHacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.