2026-11-10 · America/Los_Angeles · 社区动态 · #8

What I learned measuring an LLM agent that writes tests before the code exists

I've been building an open-source (MIT) QA workflow, Agent QA Kit. A Claude Sonnet agent, run through Claude Code, writes one unit test and one browser test per acceptance criterion, from the ticket, before the feature exists. Each test has to fail first. A person approves the tests once, then they are frozen. When the feature is done, a Claude Haiku agent runs them, and plain scripts check its verdict against the actual test reports. The agents never decide alone whether something passed. To see whether this holds up, I wrote 17 small tickets for a demo shop, each with hidden tests and a…

热度 39.5 / 100；排名与评分保留该期记录。