Deep Dive · Topic tracker

Coding agents and code review

This page follows one topic: at the top is where we stand now, updated weekly; below is everything we have published on it (daily briefs and deep dives), newest first, each linking to the original.

Last updated: Oct 24, 2026

Where we stand

AI coding agents, which read a codebase, edit files and run tests on their own, now write code faster than people can read it. Companies are responding by having machines review first and saving people's time for high-risk changes. Anthropic's Claude Code team published its approach on October 15: automated review can clear a low-risk change, any other change still needs a second person to approve it, and an agent's approval never counts as a person's (Claude blog). A preprint submitted on October 9 (TestSuiteBench) measured the other side. Code that agents write to a spec passes 82% of the reference tests, but the tests the agents write themselves often miss key combinations of inputs. The authors conclude that agents can build software they cannot verify (arXiv 2610.13065). Our August 19 deep dive measured the wait on the review side: AI-written code waited five times as long for a person to look at it, yet was reviewed faster once someone did (August 19 deep dive). The harness, the program wrapped around the model that decides how it reads files and calls tools, also changes scores. Our October 13 issue reported a commissioned test in which the same model solved 61% of tasks inside OpenAI's Codex and 41% inside Claude Code. But it was a single test, published by the party that commissioned it, with only two runs per combination (our October 13 issue). In long-document review, a neighboring field, our October 8 deep dive also found that the same models' review pass rates rose by 39 percentage points on average when the harness changed (October 8 deep dive). Our current reading: compare coding agents as a model plus its harness, because model scores alone can easily mislead. This reading rests only on one commissioned test and evaluations from a neighboring field, with no independent rerun of a controlled comparison yet, so treat it as a hypothesis.

Next reading to watch: whether the rollback and incident rates that teams publish rise after they hand review to machines.

Disclosure: we produce our analysis with help from Anthropic's models, and Claude Code is an Anthropic product.

Coding agents and code review

Where we stand

What we have published on it