Claude Code moved review before the pull request and sends humans only to risk

The Claude Code team has described a workflow in which agents catch most routine mistakes before a person sees the change. From June to September, merged pull requests in the repository more than doubled while the share that had to be reverted fell by more than 40%.

Review now starts before a pull request exists

Before opening a pull request, Claude runs tests, the build and the linter. A second agent reviews the diff with fresh context. Required CI, parallel bug hunting, a separate security pass and a merge queue follow on the pull request itself.

The team says Claude writes roughly 90% of Claude Code's code. Across Anthropic's monorepo, daily changes rose from about 2,500 in May to more than 30,000 in September. Reading every line manually therefore stopped being a credible control mechanism.

Human attention is allocated by the impact of the change

A classifier assigns one of three tiers. Documentation or copy can pass with no additional person after automated checks are green, ordinary product changes require one person and the highest-risk code requires two. Sensitive paths such as billing and CI always require a person's sign-off.

That changes the senior engineer's job. Instead of hunting routine defects, people decide on architecture, scope and risk. For teams deploying coding agents, this matters more than the choice of model: review becomes a system for routing scarce human attention.

Anthropic's own metrics are not yet a universal playbook

The figures come from the team building both the tool and the process. This is not an independent experiment, and the workflow depends on high test coverage, feature flags, gradual release and fast rollback. Without those supports, fewer human eyes would merely give defects a faster route to production.

Automated review can also manufacture work for itself. One pull request went through more than 50 review-and-fix rounds. After caps and diff-only re-review, the share needing five or more rounds fell from 8.9% to 1.5%.

Reverts, incidents and human wait time will settle the question

The useful signal is not the number of agent comments but the trend in reverts and incidents alongside the time a pull request waits for a person. Anthropic tracks the median and 95th percentile of sign-offs by people and turns past incidents into a daily eval.

Teams adopting the approach should start with one repository, define risk tiers and keep sensitive areas under human control. Only stable production data will show whether automation saved attention or merely moved the bill to the on-call engineer.

Lilith's verdict

When an agent writes 90% of the code, a person cannot stand over every line. They need to sit at the junction and spot which change is heading for the cliff.