Claude Code team uses verification loops for mostly agent-written code
Original title: How the Claude Code team thinks about verification
A separate grader raised structured-file task success by up to 10 points versus a standard prompting loop in internal tests.
The Claude Code team says roughly 90% of its code is written and reviewed by Claude, so agents need evidence-based checks before more work is delegated. End-to-end runs, property tests, and differential comparisons with an older system are the checks it prioritizes, because generated tests can run often without many extra tokens. In internal testing, a separate grader scoring output against a rubric improved success by up to 10 points, with the largest gains on the hardest problems. The post also describes a repo skill named verify for starting simulators and apps.
Unconfirmed: The roughly 90% figure and the 10-point gain are the team's own internal claims.
- The team says about 90% of its code is written and reviewed by Claude.
- Flaky tests can train an agent to rerun checks instead of fixing code.
- Property tests cited include fast-check, Hypothesis, and proptest.
- Differential tests compare old and new systems on random or production inputs.
- Vendors
- Anthropic
- Products
- Claude Code
- AI models
- Claude
From the original article
10 min read at claude.dev
On the Claude Code team, the vast majority of our code (~90%) is written and reviewed by Claude . At this scale, how we equip our agents to verify their work with evidence becomes extremely important. Verification loops are the building block that lets you delegate more work to Claude.
By craigmart · claude.dev
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.