GitHub's agents refactor its code. Humans merged 79%.

GitHub has been running autonomous agents against its own codebase to find and fix structural problems, and across the fleet those agents merge at 76 to 79 percent. The number reads like a rebuttal to everyone who says agent-written code is too rough to ship. The detail that explains it is the one nobody puts in the headline: the changes are tiny, and a person still approves every one.

A fleet, not one lucky workflow#

The work runs under Peli's Agent Factory, published on the GitHub Agentic Workflows blog on January 13, 2026 and authored by Don Syme and Peli de Halleux. Each agent targets one class of structural debt. The Semantic Function Refactor finds functions sitting in a file where they no longer belong and drove 112 merged pull requests out of 142 proposed. The Large File Simplifier picks the single largest file over a threshold each weekday and proposes a split, at 26 merged out of 33. The pattern repeats across the rest: the Duplicate Code Detector at 76 of 96, the Go Module Usage Expert at 58 of 74, the Typist at 19 of 25. GitHub Agentic Workflows entered technical preview a month later, on February 13.

Read the numbers together and the consistency is the story. Five different agents, five different structural problems, and a merge rate that sits near four in five every time.

Why the number is that high#

The tasks are narrow and mechanical by design. The Large File Simplifier files at most one issue per run and skips the day entirely if a matching issue already exists. The refactor agent closes stale issues before creating new ones so recommendations never pile up. Every proposal is one function moved, one file split, one duplicate removed.

Scope is the variable that decides whether an agent change survives review. An arXiv study of failed agentic pull requests, Where Do AI Coding Agents Fail?, found that unmerged agent contributions skew toward larger changes, higher rates of CI and test failure, duplicate submissions, and features nobody asked for. The Agent Factory design removes each of those failure modes before the pull request exists. So 79 percent is not a claim about agent autonomy. It is a claim about restraint.

The part GitHub says out loud#

GitHub is direct about the mechanism. Its February 13 writeup states that "pull requests are never merged automatically, and humans must always review and approve." Every one of those merged refactors passed a maintainer. The agent proposes the change. A person decides whether it ships.

That is the reconciliation for anyone who has seen the other set of numbers. LinearB's 2026 engineering benchmarks found that AI-assisted pull requests merge at less than half the rate of human-authored code, even as more than 88 percent of developers now use AI tools daily. Hyrax has covered the same gap in nearly half of agent PRs get rejected and in the FrontierCode merge bar. The two pictures stop contradicting each other the moment you separate scoped maintenance from broad feature work. Small and reviewed clears the bar. Large and unsupervised does not.

What it means for autonomous review#

The Agent Factory is a scheduled job. One file a weekday, an issue a team turns into a pull request, a maintainer who merges. It is also a framework a team assembles and maintains inside its own GitHub Actions. The limits are coverage and timing. Structural debt does not arrive one file per weekday. It arrives on every pull request, as a function gets appended to whatever file already exists and boundaries blur a commit at a time.

Hyrax applies the same proven shape continuously instead of on a timer. It reviews all code in a repository, whether a person or an agent wrote it, finds the structural problems as they arrive across every pull request and commit, writes the fix in an isolated worktree, runs 13 verification steps, then submits the pull request. The user merges. The four workflows are Scan, Fix, Improve, and Govern, and none of them merges anything on its own.

The 79 percent figure is the useful part of the story, and it points the opposite way from the headline fear. Agents can do real structural work that survives human review, as long as the work stays small and someone still owns the merge. The question for a team is not whether to hand an agent the whole codebase. It is how much scoped, reviewed, per-pull-request maintenance it is leaving undone.

Hyrax is live at hyrax.dev.