GitHub shipped two things to Copilot code review this week.
First one is a new review level called Max. It sits next to Lite and Balanced. It runs several review agents on models from different providers, then one more agent goes through what they found, settles the disagreements and writes one review. It's slower and costs more credits.
Second one is the bigger deal. Copilot approvals are now generally available. If an admin turns it on, Copilot's approval counts toward the required approvals on a PR. Same as a human. Push new commits and the approval gets dismissed, same as a human too.
So yes, an agent can write the PR and another agent can approve it 🤣
It sounds risky until you read the settings.
it's off by default
admins pick which file paths Copilot is allowed to approve, up to 15 globs
the approval only counts if every changed file matches one of them
admins can cap which review level is even available
That's the interesting part. It's not "trust the AI". It's "trust the AI inside the box you drew". Same pattern as agentic coding in general: the agent gets trustworthy when the constraints are written, not when the model gets smarter.
The sensible setup is docs, copy changes, tests, maybe some UI. Things where a wrong approval burns nothing. Auth and payments stay with a human. Then widen the globs slowly as the reviews prove themselves.
The multi-model part makes sense too. Different models miss different stuff, so making them check each other is what people already do by hand when they paste a diff into a second model. Now it's a dropdown.
A human reading every line of every PR was already not really true on most teams. At least now the gap has rules.
