Skip to content
Hot eventDeveloping

Anthropic's Claude Code team details how it verifies agent-written code

1 reports1 reporting sourcesUpdated 10 h ago

Start here

AI digest

Anthropic's Claude Code team says roughly 90% of its code is written and reviewed by Claude, and it relies on verification loops to delegate work at that scale. In internal testing on structured file generation, a separate grader checking Claude's output against a rubric improved task success by up to 10 points over a standard prompting loop, with the largest gains on the hardest problems, the team says. The team favors end-to-end, property and differential tests, and has Claude run a verify skill before committing changes. If a check still fails after three rounds, Claude stops and brings in a human. Checks live on a branch the change cannot modify; an observe mode logs what would have been blocked so false positives can be reviewed, and if commits are blocked twice in a session the next one is allowed through so Claude does not get stuck.

Generated from reports · updated 1 h ago

Report timeline

Follow the reports to see the event from different sides.

November 16
  1. Anthropic · Claude.devPicks
    How the Claude Code team approaches verification for agent-written code

    The Claude Code team says about 90% of its code is written and reviewed by Claude, and it relies on verification loops to delegate at that scale. In internal testing on structured file generation, a separate grader checking Claude's work against a rubric improved task success by up to 10 points over a standard prompting loop, with the largest gains on the hardest problems. The team favors end-to-end tests, property tests and differential tests, runs a verify skill before commits, sends a check to a human after three failed fix rounds, and keeps checks on a branch the change cannot modify.

Heat for this event

Not enough continuous observations yet to draw a trend.