Agents that Run While I Sleep

Why It Matters

Ensuring correctness of AI‑generated code prevents production bugs and maintains developer trust as autonomous agents scale.

Key Takeaways

• AI agents produce code faster than manual reviews
• Self‑checking AI creates a self‑congratulation loop
• Acceptance criteria act as external specification for verification
• Playwright + Claude skill automates test generation and verdicts

Pulse Analysis

The rise of large‑language‑model assistants such as Claude has turned code writing into a near‑continuous background process. Teams report merging 40‑50 pull requests per week, far outpacing the capacity of human reviewers. While the speed boost accelerates delivery, it also erodes the safety net that traditional code review provides, exposing production to subtle regressions and mis‑interpreted requirements. As autonomous agents become commonplace, organizations must replace the missing second set of eyes with systematic, model‑agnostic verification that can scale with the volume of generated changes.

Test‑driven development offers a proven remedy: define the expected behavior before any code exists, then let the model produce an implementation that must satisfy those tests. By writing clear acceptance criteria—such as exact login redirects, error messages, and rate‑limit responses—developers create an objective benchmark that the AI cannot cheat. Automated Playwright or curl checks execute each criterion, capture screenshots or response data, and return pass/fail verdicts. This approach isolates human effort to the handful of failures, turning exhaustive diff reviews into focused debugging sessions.

The author’s open‑source Claude Skill, verify, stitches together four lightweight stages: a pre‑flight bash guard, an Opus planner that maps criteria to selectors, parallel Sonnet agents that drive the browser, and a final Opus judge that synthesizes evidence into JSON verdicts. Because each stage is a single claude ‑p call, the workflow plugs directly into CI pipelines without extra infrastructure or API keys. Teams that adopt this pattern can maintain high velocity while regaining confidence that shipped features meet real‑world expectations, paving the way for broader, trustworthy AI‑augmented development.