Round 2 runs on 25 and 26 September 2026. It is entirely individual: one participant, one agent, one domain. You own the problem, the approach and the engineering.

What you build

A working agent for your chosen domain, inside a starter repository that already runs. The scaffolding is not the challenge: the reasoning, the tool use, the grounding and the behaviour under real conditions are.

Submission

Submissions close on 27 September 2026. Working code in the repository. Merged, or it did not happen.

• The agent runs from a clean checkout
• Decisions are documented where a reviewer will find them
• Failure modes are handled, not hidden
• You can explain every line you submitted

AI tooling

ChatGPT, Claude, Gemini, Copilot, Cursor and equivalents are all permitted. You remain responsible for understanding, testing, validating, explaining and defending whatever you submit. That responsibility is the actual assessment.

Evaluation

Submissions are assessed between 28 September and 5 October 2026, within each domain, so every pod formed for Round 3 gets one specialist per challenge. Finalists are announced on 6 October 2026.

Submissions close 27 September 2026.