GitHub Agent Merge: an agent that finishes the pull request

Coding agents have been able to open a pull request for a year. The tedious part was never opening it — it was the nineteen minutes after review, fixing the nit, rerunning the check that flaked, rebasing onto a branch that moved. GitHub just pointed an agent at that. Agent Merge is in public preview, and per the Copilot weekly release notes it "gets your pull request ready to merge by resolving review feedback, failed checks, and merge conflicts."

What actually happened

Agent Merge landed in the August 31 Copilot release, published September 4, as a public preview in VS Code. It is the end of a visible progression rather than a standing start. GitHub shipped @copilot-mentioned merge conflict resolution on pull requests in March, a three-click version in April, and the same capability on GitHub Mobile in July. Agent Merge bundles conflicts with the other two things that block a merge: review comments and red CI.

Three other items in the same release matter if you run agents at work:

• Multi-root workspaces went experimental, bringing Copilot and Claude agent sessions to every folder in a workspace rather than one root
• The Copilot app and CLI now honor content exclusions, so paths you have excluded stay out of agentic workflows
• Claude Fable 5.1 reached Pro+, Max, Business and Enterprise; Gemini 3.8 Flash reached Pro, Pro+, Max, Business and Enterprise
• The JetBrains Copilot harness went generally available

Why this matters for your business

Look at what Agent Merge automates and notice what is left. Writing the code: agent. Fixing the review comment: agent. Making CI green: agent. Rebasing: agent. The only remaining human step in that pipeline is deciding the change was right in the first place.

That is a real productivity win and a real concentration of risk in one act. Review has quietly been carrying two jobs — catching bad code, and being the friction that made someone look. Agent Merge removes the friction and leaves the catching. If your review culture was "approve, the tests pass," the tests passing now means an agent made them pass.

So the practical move is not to refuse the tool. It is to make the one surviving gate load-bearing. Concretely: require a human approval that cannot be satisfied by the same account that opened the PR. Protect the branch so checks are required and cannot be edited by the agent's own commits. Keep at least one check the agent cannot rewrite — a security scan or a build in a clean environment. And read the diff Agent Merge produced when it "resolved review feedback," because resolving a comment and addressing it are different verbs.

Also: it is a preview, in one editor. Turn it on for one repo where a bad merge is cheap, watch a dozen PRs, and decide from evidence rather than from the changelog.

Sources: GitHub Changelog, GitHub Changelog — merge conflicts.