The agents open the pull requests

On 28 October GitHub published Octoverse, its yearly report on activity across the platform, and one line stands out: by September 2026, more than 40 percent of pull requests on GitHub started with an agent’s first draft, up from 17 percent in March. That was 110 million pull requests in September alone. The same day, at its Universe conference, GitHub put Copilot Projects into public preview and made Agentic Workflows generally available.

A week later Meta bought Conductor, a start-up built to run several coding agents at once. Together the news points one way: writing code is no longer the scarce part of making software. Specifying it, reviewing it, testing it and answering for it are.

What GitHub counted

These are GitHub’s own numbers, measured on its own platform, and they should be read that way. Octoverse counts a pull request as agent started when its first draft came from a coding agent that GitHub can detect, and the report calls that signal “a floor, since not all agent use may be detected.” It covers public and private repositories, and the figures for earlier years were recalculated on the same basis.

The report also says agent pull requests now merge at least as often as the rest. In February, 83 percent of agent-involved pull requests merged within two weeks, against about 86 percent of those without a detected agent; by September the agent share had reached 90 percent. They are broader, too: about 87 percent change more than one file, against 79 percent of the others. In our reading, a merge shows that a reviewer accepted a change, not that the change was right.

More agents per person

The products GitHub launched the same day assume this share will keep growing. Copilot Projects gives work that spans several repositories one place for its context, agents and results. You describe a goal, Copilot runs agents in parallel, “checks in when it needs you” and brings back results such as pull requests for your review. Agentic Workflows let a team describe tasks such as issue triage or the analysis of a failed build in Markdown and have a coding agent carry them out in GitHub Actions, with read-only permissions by default and a separate job that scans proposed changes before they are applied.

GitHub’s Universe post describes the division of labour in one sentence: “Developers set the direction, make decisions, and stay accountable for what ships.” An analyst quoted by InfoWorld spelled out what that means in practice. “Ten parallel agent sessions, for instance, could generate ten pull requests that need to be reviewed,” said Amit Kumar Jena, who leads AI development at the consulting firm Kanerika.

Seventy percent is not done

A new test outside GitHub shows why that review cannot be skipped. On 26 October researchers from Alibaba and several universities published Co-Design Bench: 273 tasks drawn from 26'314 real user coding sessions, each checked against hidden acceptance criteria written by experts, 2'331 in all. The strongest of ten models, run in the same agent setup, met 69.9 percent of the criteria per task on average, but fully satisfied only 26.4 percent of tasks. The Decoder, reporting the study on 8 November, wrote that coding agents still “need human cleanup on nearly three out of four tasks”.

The study has limits. Each task is one fixed request, so the agents could not ask questions or be corrected along the way, the grading is done by AI judges that can be wrong, and several authors come from Alibaba, whose own Qwen model was among those tested. Still, the gap it measures, between most of the requirements and all of them, is exactly the gap a reviewer is there to close.

What good teams do now

None of this needs new tools as much as old discipline, applied more strictly. Start with the specification: when an agent drafts the change, the written task is the real source, and Jena warned that “poorly defined tasks could lead to repeated execution, rework, and higher AI credit consumption.” Write down what done means before the agent starts, the way Co-Design Bench sets its acceptance criteria first.

Keep changes small enough to read. GitHub says stacked pull requests, which split related work into smaller changes that are reviewed in sequence, had been used by more than 264'000 developers by September, two months after their preview launched. Make tests the gate: a change should not reach a person until it passes them, and the tests themselves deserve the closest review, because an agent that may edit its own tests can make anything pass.

Finally, name the person. Every change that ships needs someone who approved it and can explain it, and that someone cannot be the agent. Jena’s advice to CIOs is the same: “require human approval before agent-generated pull requests are merged.”

Tools can be switched off

The Conductor deal adds a quieter lesson. TechCrunch reported on 4 November that Meta had acquired the start-up, founded in 2024, whose agent teams are “designed to help software engineers work on separate, compartmentalized parts of a project in concert”, using agents from labs such as Anthropic and OpenAI. Meta confirmed the purchase but gave no terms.

Conductor’s own notice sets the timetable. New sign-ups stopped on 4 November, its cloud workspaces go read-only on 13 November, and on 20 November the company begins deleting cloud user data, so customers have sixteen days to export what they keep there. The desktop app keeps working for now and starts winding down in early 2027. Conductor says it has never sold or trained on user data, and it offers an export and a transfer to the local app.

Nothing here suggests bad faith. It shows how new and unsettled the layer that runs agents still is: a service that holds your agents’ workspaces and chats can be bought and closed within weeks. Keep what matters in places you control: the code and its history in your own repository, specifications and decisions written down next to it, and tests that any agent must pass, whichever vendor supplies it.

Where we stand

Werk, our app platform, is live and built on this division of labour. AI writes most of the code on a framework we built and trust, people decide what to build, design how it feels and review every release, and every release is tested. The code is yours, and the platform keeps growing under it. The next step we are working towards is agents that ship routine changes and fixes on their own, tested and reviewed before release, which only works when the specifying, testing and reviewing are already in place.