Stop being the code review bottleneck

Code review used to be the thing that kept engineers honest. Now it's the thing that's about to break them, because AI agents can write pull requests faster than any team of humans can read them. PostHog's answer isn't to speed up reviewing. It's to mostly stop doing it — by building pipelines where agents check agents, and humans only get pulled in when something actually smells wrong.

The core rule from PostHog's engineers is simple but easy to violate: the agent that wrote the code cannot be the one grading it. Self-review is a blind spot problem, not a competence one. So engineer Paul D'Ambra runs a setup called qa-swarm that spins up four separate reviewer agents — one hunting security holes like SQL and prompt injection, one covering rollout and naming conventions in his own voice, one applying Extreme Programming principles, and a broader technical team scanning for performance and database issues. A second step, review-triage, sorts every comment thread into three buckets: fix it automatically, leave a minor note, or kick it up to Paul because it's genuinely ambiguous. The loop repeats up to three times until nothing new gets flagged. It's not cheap — D'Ambra estimates 60% of his token spend goes to this kind of CI and review automation — but he says he doesn't regret a dollar of it.

Then there's the approval bottleneck nobody talks about: the actual GitHub stamp. PostHog used to run a Slack channel, #dev-stamp-exchange, where engineers begged colleagues for a quick rubber-stamp approval on low-risk changes. It worked, but it also meant constantly yanking someone out of flow state to bless code they had zero context on. Their fix, an agent called StampHog, now handles roughly one in three merged PRs on its own, checking for merge conflicts, risky keywords like auth or billing, diff size under 500 lines, and a basic LLM sanity pass. Last month alone it cleared 1,600 PRs without a single Slack ping. Anything risky still gets routed to a human, based on code ownership and git-blame history.

The more interesting philosophical shift comes from engineer Daniel Visca, who argues you shouldn't trust an agent's explanation of why its code works at all — because a wrong explanation can sound exactly as confident as a right one. His fix is to force agents to break big changes into small, independently runnable pull requests, stacked with a tool called Graphite, so each layer can be observed running before the next one gets built on top of it. Watching code behave beats reading an agent reason about it, especially past a few thousand lines. It also means smaller diffs are exactly the kind StampHog can auto-clear, which quietly closes the loop between the two ideas.

What's happening at PostHog isn't really about review speed. It's about redesigning the org chart of trust — deciding in advance which decisions a machine can make alone, which ones need a second machine's opinion, and which ones still require a human with actual context. Teams that skip that redesign are the ones that will drown in PRs they can no longer keep up with.