Since agents started opening pull requests, most teams have the same complaint: the queue is full and nobody trusts the review.
Reviewing agent output is a different job from reviewing a colleague's work, because the failure modes are different. Here is a process shaped around that.
Start with the dependency list, not the diff
A human contributor rarely adds a package without mentioning it. An agent adds three and says nothing.
GitHub hardened two supply-chain defaults on 28 July for exactly this reason — Dependabot now alerts on malicious packages across more ecosystems, and Actions holds suspicious workflows for approval. Those defaults are a safety net, not a review. Read the lockfile diff first, every time.
Check the change against the request, not against itself
Agent code is internally consistent almost by construction. It compiles, it is stylistically uniform, and it reads well. That is why reviewing it line by line feels productive and catches so little.
The common failure is not a broken line. It is a change that solves a slightly different problem than the one asked for. Re-read the original issue before the diff.
Read the tests before the implementation
Tests written alongside an implementation tend to encode what the implementation does rather than what it should do. Read them first, while you still have your own idea of correct behaviour in mind, and ask whether these tests would fail if the code were wrong.
Make the change small enough to review
GitHub put stacked pull requests into public preview on 30 July. This is the reason.
A single agent pass produces migration, model, handler and tests together. Split it into a stack and each step gets a real review instead of an approval.
Use the automated reviewer as a first pass, not a verdict
Copilot code review reached GA with agent skills and MCP on 29 July, which lets it check a diff against your own systems and conventions. That catches a class of issue no human reviewer reliably catches.
It does not catch the change that solves the wrong problem. Keep a person on that.