Skip to content

Build1 publisher3 min readPublished

Since 1 September a Copilot approval can satisfy GitHub's required-approvals rule

GitHub put Copilot approvals into public preview on 1 September 2026. The toggles that decide whether an approval counts live in Copilot settings across three scopes, not in the rulesets a policy-as-code audit reads.

The Engineer · Build desk

Illustration accompanying Since 1 September a Copilot approval can satisfy GitHub's required-approvals rule

What happened

  • GitHub put Copilot approvals into public preview on 1 September 2026, letting administrators authorise Copilot to submit an approving review that counts toward a repository's required-approvals rule.
  • Every Copilot review now states whether Copilot considers the pull request ready to approve, and GitHub's documentation says that assessment on its own does not count toward merge requirements.
  • A follow-up on 11 September gave reviews the Copilot SDK's full shell tools behind the Copilot agent firewall and an ensemble of agents at the Lite effort level.
  • The approval settings are documented in enterprise, organisation and repository Copilot pages, and none of them is configurable through rulesets, which only govern automatic review requests.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

  • decision Whoever holds enterprise AI controls now decides whether a repository's approval count can be met with one fewer human, so the team that owns the merge policy and the team that owns the toggle may not be the same people.
  • constraint Shops that verify branch protection by diffing ruleset JSON in CI have no line to assert on here, because the approval settings are outside the rulesets API that the check reads.
  • exposure Any identity that can push to a branch and re-request a review can now assemble the approval it needs without a colleague. Repositories whose rule is a single approval are the ones exposed.
  • capability The reviewer can build and test the code it reviews before it approves, which puts the Copilot agent firewall configuration on the path to a merge rather than beside it.

Set the enterprise switch, then the organisation one, and a Copilot review satisfies a required-approvals rule "the same way a teammate's approval would", according to GitHub's changelog [5]. Push a new commit and the approval is dismissed exactly as a human's would be, and the review can be re-requested [6]. On a repository whose rule is one approval, the whole path is: author pushes, Copilot approves, merge [21].

The settings live on three separate pages [19]. At enterprise level the path is AI controls > Available Agents > Copilot code review, then Allow Copilot to approve pull requests [11]. At organisation level it is Settings > Copilot > Code review > Approvals > Count Copilot approvals toward merge requirements [12]. At repository level it is Settings > Copilot > Code review > Auto-approval [13]. The repository level splits the question in two, so Copilot can post an Approve review that shows in the timeline without that review satisfying branch protection [14].

Approvals are not configured in rulesets. Automatic review is, as the rule Automatically request Copilot code review with sub-options Review new pushes and Review draft pull requests [15], exposed in the REST rulesets API as rule type `copilot_code_review` with parameters `review_on_push` and `review_draft_pull_requests` [16]. A CI job that diffs ruleset JSON can therefore tell you whether review is requested on every push and cannot tell you whether the resulting approval counts toward the merge requirement [15].

GitHub's figures for the 11 September ensemble are 47% more addressed comments per review for high severity findings, 31% for medium, 11% for low, and about 8% less review cost [10]. The high-severity gain is roughly four times the low-severity one [18], so the number that transfers to your repositories depends on the severity mix of findings you actually get, at the Lite effort level GitHub tested [9]. The metric counts comments marked addressed, and the resolution reasons that produce that label (Addressed, Won't fix, Incorrect) shipped two weeks earlier, on 27 August, alongside the removal of the 300 file and 20,000 line cap [8]. The same 11 September release also let Copilot resolve its own comments when a later commit addresses them [9]. The changelog does not say whether those self-resolutions count as addressed [9].

The dev.to post that catalogues the settings said, "This is the moment where AI code review stops being a suggestion layer and becomes part of the control plane" [17]. I would draw the line more narrowly. The merge semantics change only where the organisation setting Count Copilot approvals toward merge requirements is on [12], and GitHub is explicit in the same release that "An approval assessment alone does not count toward merge requirements" [4]. Until August, a Copilot review left only a Comment and the documentation said it did not count toward required approvals [1]. Copilot approvals are a public preview, and GitHub's docs say they are "subject to change" [7].

What to watch

  • Whether GitHub moves approvals out of public preview and into the rulesets API, where a policy-as-code diff would catch them.
  • Whether GitHub publishes how the enterprise, organisation and repository settings compose on a single documentation page.
  • Whether the ensemble's severity-weighted gains hold once Copilot's own auto-resolved comments are separated from ones humans marked addressed.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories