Trust the output, sheet 02 of 13
Review, gating and provenance for pull requests that agents open, so nothing approves itself into production.
Branch protection, CODEOWNERS and one reviewer whose approval counts. Start with warnings, then turn them into blocks.
← → move to the next sheet over, Esc zooms out.
An agent can open a pull request on its own now: the branch, the code, the tests, even a tidy description. Somebody, or something, still decides whether it goes in. That decision is the door before the merge, and in a lot of repos it's held shut by habit, not by a lock.
Think of the pass in a restaurant kitchen, the shelf where finished plates wait for the head chef. Agents are very fast line cooks. The pass is where someone looks at the plate before it leaves, and the kitchen that skips it hears about it from the customer.
Review tools have got good at commenting. Very few are allowed to say yes or no, and the ones that can ship with that switched off. So the door is still something you assemble: branch rules, code owners, one reviewer whose approval counts, and a written rule for which changes a person must see.
Every link opened on 19 Nov 2026. The numbers match the squares on the map.
Built in
Branch rules that block a merge until conditions are met, and a file that names who owns which paths. The frame every other tool on this sheet hangs from.
The catch. They enforce whatever you configure. Leave them at the defaults and the door is open.
Product, GitHub
Copilot reviews a pull request and suggests fixes. With Copilot approvals enabled, its approving review can satisfy a required-approval rule the way a teammate's would.
The catch. Its reviews don't count toward required approvals by default. Turn that on deliberately, and limit where it applies.
Product, Anthropic
Several agents read each change against the surrounding code, verify their findings and comment inline, ranked by severity. For Team and Enterprise plans.
The catch. A review never approves or blocks the pull request. The door is still yours to build.
Product
AI review that reads a pull request with the context of the whole codebase, with review tiers and an option to auto-approve pull requests.
The catch. Auto-approval is a setting you own. Decide the paths it may never touch before you turn it on.
Product
AI code review plus triage that ranks open pull requests by risk and urgency, so a person knows which to read first.
The catch. Triage tells you where to look. It doesn't decide what merges.
Product, free and Enterprise tiers
A proxy between your package managers and the registries that checks each package as it's requested and blocks malicious ones at any depth, transitive dependencies included.
The catch. It guards the install, not the code the agent wrote around it.
Open source, MIT
Agents in GitHub Actions that hold no secrets and can only stage their writes. A separate job filters, caps and cleans each write before anything reaches the repo.
The catch. Its own site still says to review the changes before you merge them, and its authors call zero-secret agents a trade-off between security and utility.