Chinese
Uncategorized

GitHub’s Agent-Written Pull Requests Move the Bottleneck to Review

GitHub’s 2026 Octoverse report shows agents moving into the start of the PR flow. The hard part is now review.

Jacky Wang 4 min read 92 views

Last week, while changing a small feature, I noticed that writing the files was no longer the slow part. An agent can open several modules, add tests, run commands, and leave behind a pull request that looks tidy. The expensive part is deciding whether the change stayed inside the intended boundary. I need to know what it changed, why it changed it, which checks actually exercised the risk, and which assumptions are still mine to approve.

GitHub’s 2026 Octoverse report puts numbers around that shift. GitHub says that more than 40% of pull requests started with an agent’s first draft by September 2026, up from 17% in March. It estimates 110 million agent-started pull requests in September alone, roughly six times the March volume. That is not evidence that agents have replaced software engineers. It is evidence that code generation is moving earlier in the engineering workflow.

More generated code means a wider review surface

The report also says agent-involved pull requests were more likely to change multiple files: 87%, compared with 79% for pull requests without detected agent involvement. Multi-file changes are often the right answer. A feature may need a route, a database query, a UI state, and a test. The problem is that the review cost becomes a graph problem rather than a line-count problem.

One small configuration change can alter permissions. One migration can change what an old worker reads. One retry branch can turn an external request into a duplicate charge. A green test command does not make those relationships obvious.

I prefer to make the pull request a verifiable change package. A minimal local gate might look like this:

git diff --stat origin/main...HEAD
pnpm test
pnpm typecheck
git diff --check origin/main...HEAD

Those commands only catch cheap and visible failures. They do not establish correctness. The agent should also state the observed behavior, list the paths it touched, name tests it did not run, and call out assumptions that need a human decision. Without that record, the reviewer has to reconstruct the task and guess what the agent believed it was doing.

GitHub reports that about 90% of agent-involved pull requests were merged within two weeks in September, compared with about 86% for the comparison group. I would not read that gap as a quality score. It could reflect smaller changes, different teams, or stronger automation. GitHub explicitly frames the figures as observed usage rather than model quality, and that distinction matters.

Review the failure paths, not the confidence of the prose

When I review agent-written code, I look first at the parts an agent cannot easily prove by saying “tests pass”: failure paths, authorization boundaries, data compatibility, retry behavior, and secret handling. A webhook handler can pass its happy-path test and still create two orders when the provider retries. A type-safe Kubernetes manifest can still grant a ServiceAccount more access than the service needs.

That is why I would rather receive a compact verification table than a polished paragraph. For a new endpoint: normal request, missing identity, and duplicate request. For a schema change: old data remains readable. For an upstream timeout: the worker does not charge twice. These statements give the reviewer something concrete to check and give the next engineer a record of what was promised.

The Octoverse report says new repositories with an agent configuration file averaged eight times as many commits as new repositories without one. That is an interesting signal, not a guaranteed effect of writing AGENTS.md. Active teams may simply be more likely to create a configuration file. I read it as evidence that some teams are treating working rules as repository infrastructure.

My merge boundary stays human

I will give agents more implementation work, but I will keep merge authority separate. An agent can create a branch, run tests, open a pull request, and respond to failure logs. It should not expand its own permissions, change production secrets, or merge a change whose acceptance criteria are still vague.

The useful unit is not “code produced by AI”. It is a traceable engineering proposal: reproducible, explainable, and reversible. Generation will get faster. My preference is to spend ten extra minutes reviewing a risky multi-file diff rather than discover on a weekend that a tidy pull request quietly changed how production behaves.

Sources: GitHub 2026 Octoverse, GitHub agent apps

发表评论