9pm before the deploy: reading a list of findings, triaging severity, writing fixes for the three that matter, deferring the rest until they’re forgotten.
The comment didn’t do the work. It added work.
What autonomous means in code review
An autonomous agent reads your recent commits with the same fresh-reader advantage a human reviewer has: no context about what you intended, only what the code actually does.
It finds the middleware that strips the auth header before the downstream service checks it. The cache that returns stale data after a failed write never got invalidated. The test that passes because the mock never actually covered the path you changed last week.
Then it writes the fix and commits it.
You don’t triage findings. You review a diff. The bottleneck shifts from “find the problem and fix it” to “decide if this fix belongs.”
The actual workflow
Agents run against your codebase while you’re offline. Each one reads recent changes, applies checks for failure patterns, and submits fixes as commits. You review what shipped in the morning.
No session to maintain. No comment thread to manage. Output is git history.
What it doesn’t replace
An autonomous agent doesn’t catch architectural decisions made in the wrong direction. It doesn’t know your product well enough to flag features that work correctly but solve the wrong problem. It doesn’t replace the judgment call about which technical debt is load-bearing.
Those reviews still belong to you. What it replaces is the review that generates a finding you already know how to fix but haven’t had time to action.
What actually accumulates
The real cost of skipped code review isn’t any single bug. It’s the pattern. Null checks that aren’t there. Error paths never wired. Test coverage stuck at “good enough.” Each deferred review is a small bet that the edge case won’t fire before you get back to it.
Most of the time the bet pays off. When it doesn’t, you spend a weekend on a bug a reviewer would have caught in ten minutes.
Autonomous review doesn’t require a reviewer. It requires a written spec and a git repo.
The problem wasn’t finding the bugs. It was finding time to fix them.