GitHub Agent Merge: Watch and Fix Pull Requests

What changed on October 28

The October 28, 2026 announcement integrates Agent Merge into pull requests on GitHub. It was already available in the GitHub Copilot app and Visual Studio Code. The new entry point is Watch and fix in the PR merge box, with separate choices for failed Actions workflows and merge conflicts.

For a first evaluation, treat the repair loop as the product you are testing. Ask whether the watcher notices the intended blocker, makes an understandable correction, and leaves evidence you can review. Decide what a correct repair means before inviting an agent to make one.

Check eligibility and the person enabling it

GitHub's operating guide requires paid Copilot access to Copilot in the cloud, applicable organization or enterprise permission, and the repository's Enable automations policy. Fork PRs and stacked PRs are unsupported. AI credits are charged to whoever enables the watcher, which may be someone other than the author.

Before the pilot, write down the repository, branch, author, and person who will enable it. Have the repository owner check the policy and have the enabling user confirm the usage arrangement. If the control is missing, compare that inventory with the documented prerequisites instead of repeatedly changing the test PR. Keep billing ownership in the trial notes so a teammate does not mistake authorship for responsibility for usage.

Choose one repair mode on a disposable PR

The following is a proposed reader test; we have not executed it. Use a same-repository, non-stacked PR in a disposable repository with a small, understandable application and a working test suite. Leave auto-merge off for the trial. Record the starting commit and passing workflow result, then introduce a small implementation mistake that an existing assertion detects. For example, change a function that should return a trimmed name so it returns the original string, while retaining the assertion for surrounding whitespace.

In the merge box, open Watch and fix and select Fix failed Actions workflows for this trial. Keep conflict repair outside this first experiment so the reason for each agent action is easier to follow. Preserve the failing run's link and the expected behavior. Do not delete assertions, lower coverage thresholds, skip the failing job, or ask the agent to weaken tests to obtain a green result.

For a separate conflict trial, start from a fresh disposable PR and select Resolve merge conflicts instead. Create a deliberate disagreement between the PR and base branch over a small function whose expected behavior you have written down. Choose a case where the correct combined behavior is clear to a reviewer. Avoid beginning with a complex refactor: that would make it harder to tell whether a plausible-looking resolution preserved both branches' intent.

Follow the repair, then judge the current result

GitHub documents merge-box updates and a timeline event explaining why work started. Repairs run in an Actions-based development environment using repository setup steps, custom instructions, and agent skills. Check your repository's .github/workflows/copilot-setup-steps.yml before the trial so you know what setup you are asking the repair to use.

Capture the trigger event and compare it with the failure you introduced. Then review every changed file against your written expectation. In the whitespace example, look for a correction to the implementation and verify that the assertion remains meaningful. Investigate any edit to workflow configuration or test expectations before accepting the result. A repair that removes the signal you wanted to observe should fail your evaluation even if the dashboard becomes green.

Check the workflow results associated with the repaired revision, rather than relying on the earlier run you saved as a baseline. For the conflict trial, read the resolved function against both branches and rerun the relevant behavior checks. If you need to verify who supplied a check, the reusable-workflow identity report covers that separate provenance question. Keep the repair diff, current commit, check links, and your review decision together in the trial record.

Stop the watcher before expanding the trial

To disable it, click Unassign Copilot in the merge box; the enabling user or someone with repository write access can do this. GitHub also documents an automatic default under profile Settings, Copilot, Cloud, with separate repair switches. Keep that default outside the initial trial so your experiment stays attached to the PR you chose.

Record when you disabled the watcher and inspect the PR's displayed state. Close the disposable PR when the review is complete, and retain the evidence with a short decision: adopt the tested mode, revise the setup and repeat, or stop. Do not infer from this disable action that work already underway was canceled; the cited guide documents disabling the watcher without establishing that cancellation behavior.

Keep approval and merging as separate decisions

Agent Merge does not enable auto-merge. The latter waits for required reviews and status checks; repair still needs human review. Use the independent-approval report for approval policy, and the merge-readiness checklist for your final decision. A successful pilot should tell you whether to delegate a repair mode, with evidence that the correction preserved the intended behavior.