I'm Building Agents That Run While I Sleep

When AI agents write code autonomously overnight, verifying correctness becomes the core challenge. The author argues that having the same AI write both code and tests creates a self-congratulation loop that misses original misunderstandings. The solution draws from TDD: write acceptance criteria before prompting the agent, then run automated verification against those criteria. A practical implementation using Claude Code's headless mode and Playwright MCP is described, with four stages: pre-flight checks, an Opus planner that reads specs and changed files, parallel Sonnet browser agents per acceptance criterion, and a final Opus judge that returns pass/fail verdicts. The workflow shifts review effort from reading diffs to investigating failures, and the author open-sourced the tool as a Claude Code plugin.