Track 01
Build a factory: a team of agents that turns a one-page spec into working software, with no human writing the code.
At 14:00 your code is frozen at the commit you submitted. We then publish one or two specs nobody has seen. Run your factory on them, unchanged, and push what it produces by 16:30. Our acceptance tests run against it, and the results are shown at the demos. You may watch the run; every time a person steps in, note it in the log.
| Criterion | Points |
|---|---|
| Hidden task: our tests pass, with the fewest human steps | 35 |
| Multi-agent architecture: roles, agents working in parallel, shared state | 25 |
| Autonomy: it notices its own mistakes and recovers | 15 |
| Transparency: a trace of the run, with time and cost per task | 15 |
| README: someone else can run it | 10 |