Claude Code's sandbox, Code Mode and Projects
Plus Claude Sheets against Excel and Google Sheets
Tools
Claude blog · Addy Osmani · Nov 16
Written with AI assistance, reviewed by the ACN desk.
“Verification loops are the building block that lets you delegate more work to Claude.”
The Claude Code team at Anthropic has described how it verifies the work its agents do. The post on the Claude blog, by Addy Osmani, was published on November 16. It says Claude writes and reviews about 90% of the team’s code.
The team says verification loops let it delegate more work to Claude. Without them, the post says, a team can only run one or two Claudes at a time. With them, it says, teams can run dozens or hundreds in parallel. In internal testing, a separate grader checking work against a rubric improved task success by up to 10 points.
The post names 3 test types worth investing in: end-to-end, property and differential tests. It says running the app end to end matters more than tests. Claude tries to fix failing checks and brings in a person after 3 failed rounds. Code owners write their own checks. A change cannot edit the check that judges it.
Commits are stamped with how they were verified, including “none.” A gate can block unverified commits, with off, observe and enforce modes. The team also keeps a reliability checklist of 39 items.
“Verification loops are the building block that lets you delegate more work to Claude.”
Sponsor
Spot an error? Email corrections@aicodingnews.si. Fixes are logged on the corrections page.
Sponsor
Sponsor
Plus Claude Sheets against Excel and Google Sheets
Version 2.1.323 · experimental, opt-in
How the Claude Code team verifies agent work
Shares close up 17% after topping $2 trillion at the peak
Sponsor
Sponsor