How the Claude Code team thinks about verification
- ID
- 48038
- Status
- summarized
- Published
- 16 Nov 2026, 8:00 AM
- Fetched
- 19 Nov 2026, 6:12 AM
- Provider
- Claude
- Category
- ai-labs
- Original URL
- https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/
- Source URL
- https://raw.githubusercontent.com/leontloveless/ai-rss-feeds/main/feeds/claude.xml
Summary
- Score
- 8.0
- Created
- 19 Nov 2026, 6:13 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_users
What happened
Addy Osmani's post on the claude.dev blog describes how the Claude Code team verifies agent-written code: roughly 90% of their code is written and reviewed by Claude, and a separate grader checking output against a rubric improved task success by up to 10 points over a standard prompting loop, with the biggest gains on the hardest problems. The piece argues verification loops are what let a team run dozens or hundreds of agents in parallel instead of one or two, and recommends investing in end-to-end, property, and differential tests while avoiding flaky or slow suites. It also describes closing the loop by turning recurring agent failures into hooks or explicit CLAUDE.md rules, e.g. a test that checks logs for user data.
Why it matters
If you delegate to coding agents, the concrete lever here is a separate grader with a rubric (reported up to +10 points task success) plus encoding each repeated failure as a hook or CLAUDE.md rule rather than re-prompting. The post also warns that flaky tests teach the agent to re-run instead of fixing code, so test reliability is a prerequisite before scaling parallelism.
Discussion angle
Would a separate grader with a written rubric fit your workflow, and which repeated agent failure in your repo should become a CLAUDE.md rule or hook first?