How Claude Code Team Approaches Verification at Scale

Original: How the Claude Code team thinks about verification

Why This Matters

Verification architecture directly determines how many AI coding agents can run safely in parallel.

Anthropic's Claude Code team, where ~90% of code is written by Claude, published a framework on Nov 16, 2026 for using E2E, property, and differential tests to let coding agents verify their own work with evidence.

With roughly 90% of Claude Code's codebase written and reviewed by Claude itself, the team has developed a verification-first philosophy to maintain quality at scale. The core argument: without automated verification loops, engineers can run only one or two agents in parallel because they must manually check outputs. With solid verification, dozens or hundreds can run simultaneously.

Internal testing backs this up. A separate grader checking Claude's work against a rubric improved task success by up to 10 percentage points over standard prompting, with the largest gains on the hardest problems.

The team highlights three test types worth prioritizing. End-to-end tests top the list because they're hard to fool—either the full user flow works or it doesn't. Unlike unit tests, which can pass while a feature is still broken, E2E tests provide concrete failure evidence including screenshots. Property tests and differential tests round out the recommended stack.

Flakiness and speed are flagged as the two biggest pitfalls. Flaky tests train agents to retry rather than fix; slow tests reduce batch coverage. The team also emphasizes closing the loop: when an agent repeatedly fails a specific check, the fix goes into CLAUDE.md as an explicit rule, preventing recurrence without burning additional tokens.

Source

claude.dev — Read original →