Copilot’s monthly pull request report from GitHub’s core repo

At GitHub, the team builds github.com on GitHub itself, using the same tools shipped to developers. As Copilot has moved beyond autocomplete into a full AI assistant, the internal workflow has changed: @Copilot is now assigned issues directly, opens pull requests, and carries out the work in the core repository.

An analysis of one month’s worth of Copilot-created pull requests in that repo shows a clear pattern. Copilot is handling a broad mix of work, from quick UI fixes to high-stakes migrations, and it is doing so at a level of autonomy that goes well beyond code suggestion.

Taking on the tedious, time-consuming work

A large share of Copilot’s output is the small but constant work that adds up across a codebase. This includes fixing minor UI issues like misaligned icons and placeholder text, as well as housekeeping tasks. In one pull request, Copilot was assigned to correct 161 typos spread across comments and documentation strings in 100 files — tedious work a human engineer is glad to avoid.

Maintenance and large-scale refactors

Copilot is also assigned cleanup and modernization tasks that would otherwise eat up significant engineering time:

  • Feature flag removal: Copilot routinely cleans up deprecated flags, removing conditional logic, stale code, and outdated tests across the repository.
  • Repository-wide renaming: It has authored pull requests to rename core internal classes used across the app — a task that is slow and error-prone for humans.
  • Performance fixes: Copilot identifies and replaces inefficient patterns in high-traffic code paths.

Bug fixes and CI stability

Beyond upkeep, Copilot is actively patching production issues. It has resolved NoMethodError defects in core logic and even tackled error masking problems in caching infrastructure. One particularly impactful pull request fixed a severe performance bug where git push took roughly 15 minutes for engineers in Codespaces. Copilot is also regularly assigned to investigate and fix flaky tests that break builds.

Shipping new features and internal tooling

Copilot is not limited to legacy code. Engineers write a spec in an issue, assign it to @Copilot, and the PR comes back. This has produced new REST API endpoints, such as one for listing repository security advisory comments. Copilot has also been heavily used to build and maintain internal systems, including the intranet, training sites, and onboarding material.

Migrations, security, and documentation

Copilot is trusted with high-precision work as well. It has added security gates so internal integrations cannot modify releases or release assets, and it has handled database schema migrations such as converting column types to newer standards. It also keeps docs in sync with code — for instance, authoring pull requests that add comments to rate-limiting code so developers remember to keep it aligned with another service.

Auditing the codebase and reporting findings

Some of Copilot’s most sophisticated assignments treat it as a researcher. Given an ambiguous task, it explores the codebase and returns findings inside a pull request. In one such task, it audited every Codespaces feature flag and produced a report categorizing each flag with its references. Another task asked for a “comprehensive analysis of authorization queries,” which surfaced opportunities for performance and safety improvements. That moves Copilot from generating code toward systems-level analysis, letting engineers start at the difficult part instead of spending days on discovery.

The collaboration pattern: first pass, then human judgment

Copilot’s merge rate in a repository will generally be lower than a human contributor’s, and that is the point. The workflow is straightforward: assign @Copilot an issue, it opens a pull request with a first-pass solution, and a human reviews it like any other PR. The reviewer decides whether to merge, iterate on the branch, or close it and go another way.

The value is not in blind merges. It is in having a concrete implementation to critique instead of an empty editor. The scaffolding and boilerplate are already handled, leaving engineers to focus on the core problem, refine suggestions, and own areas where AI is not a fit—architecture, security, and user experience. Copilot is doing the tedious 80%; engineers reserve their attention for the critical 20%.