An Anthropic blog post describes how the Claude Code team approaches verification for coding agents. The core idea: agent-written code should be validated by evidence, not by the agent's own claims, using end-to-end tests, property-based tests, and differential tests. The team also runs large numbers of verifications in parallel and treats test failures as a signal for routing which changes need human review.
The piece emphasizes keeping verifiers honest — using per-change verification via skill files and tooling such as browser automation to check whether changes introduce regressions — since an agent that grades its own work can just as easily grade it wrong.
For developers using agents daily, the takeaway is practical: design your test setup so the agent has objective, reproducible checks to run, and make a conscious, project-level decision about which failures surface to a human, since no one reads every diff.
Discussion on Hacker News pushed back on the economics: commenters argued that many described practices are token-heavy compared to deterministic alternatives like pre-written Playwright suites, and that expensive per-change verification isn't realistic at typical API budgets.
Summary by angular.love. The original source has the full details.