Coding agents get benchmarked on toys: to-do apps, Twitter clones, one-page games.
I run a private security company in Italy and I build our software with AI agents. It has customers who pay for it. So every time a new model or agent shows up, I give it the same real feature inside that codebase and publish what happened: time, attempts, cost, what broke, what shipped.
No proprietary code is published here — times, behaviour and mistakes are.
| # | Task | Agents | Result |
|---|---|---|---|
| 01 | Minimum rest between shifts | Claude Code · Grok · Codex | all passed 15/15 — one found the rule already existed |
Run by @GNesca.