Agent-written code survives review as well as yours does.

The disposable-code story is wrong

The industry story about AI-written code is that it is disposable: quick to write, quick to throw away. A preprint posted this week tests that story against 2,794 merged pull requests in open-source projects and finds it does not hold. AI-authored code was not modified sooner than human code. Code from supervised assistants such as Copilot and Codex survived longer than human-written code, with a hazard ratio of 0.76. And when the code was changed, the intent behind the change was statistically similar to changes made to human code, with a Cramer's V of 0.080 — a small number that means the reasons for touching the code look the same either way. m:…0109

Adoption is not a straight line, though. A second preprint introduces GHAgentCI, a dataset of more than 15,000 workflow histories and 205,000 agent-evidence records from more than 6,000 GitHub repositories, built to track coding agents inside CI pipelines. Its exploratory analysis finds a trend that cuts against the survival numbers: agentic workflows are frequently removed from CI histories. Teams try agents in the pipeline, then pull them out. The dataset's extractor, cai-miner, identifies agentic workflows with 96.1 to 98.8 percent accuracy, so the removals are visible rather than hidden. Code that survives is one question; workflows that stick is another. m:…9309

Review got faster. Ownership got formal.

Where agents clearly help is the review queue. A third preprint quantifies the effect on GitHub pull requests: GPT-assisted pull requests cut median resolution time by more than 60 percent, from 23 hours to 9 hours, with review time down 33 percent and waiting time before acceptance down 87 percent. Developers mostly use the models for code optimisation, at 60 percent of uses, followed by bug fixing at 26 percent and documentation at 12 percent. The bottleneck the agents relieve is waiting, not thinking. m:…8938

As agents write more of the code, ownership is being formalised. A fourth preprint, covering more than 844,000 pull requests, looks at GitHub's CODEOWNERS feature, which assigns named reviewers to parts of a codebase. Adoption correlates with a reduction or slower growth in merge time and review effort, and designated owners measurably increase their review, approval and merge activity right after adoption — while contribution-based owners, the people who used to review because they wrote the code, do less review work per pull request afterwards. The review job is moving from whoever wrote it to whoever is named responsible for it, which is exactly the shift an agent-written codebase needs. m:…0108

The new bottleneck is detection

The last piece of the week is about what happens when agents fix the pipeline itself. GHA-SmellFix, an agentic pipeline for detecting and refactoring smells in GitHub Actions workflows, covers 19 distinct workflow smells with a curated ground truth of 1,710 workflow-smell pairs. Its finding is a caution: splitting the agent into a decomposed architecture does not universally improve patch quality, though it does give better control over when to attempt a refactoring. Detection accuracy is the primary bottleneck, and false positives can trigger unnecessary and potentially harmful edits to execution-relevant constructs. Put the week's papers together and the shape of the job is clear. Writing code is no longer the constraint. Knowing what the code does, who owns it, and whether the fix is real — that is the work now. m:…9303