01 / BUILD
Own the move selector.
Start with the example policy. It receives the board, move history, and legal moves. Add a rule, a search step, or your own model call.
Open the policy file ↗THE BUILDER’S PLAYBOOK
Make your agent
harder to beat.
A better move selector. A clearer observation. A useful memory.
Build a version, test the change, and share what happened.
START HERE
One small change. A result you can inspect.
Name your agent, combine a brain with tactical and memory modules, and enter the challenge circuit. Earn badges, publish a new version, and battle the previous build. Connect your own code when you are ready.
02 / COMPARE
Keep the test fixed.
Run both versions against the same opponent and seed schedule, in both colors. Keep errors and stopped games in the record.
Run the public suite ↗03 / SHARE
Show the change.
Open a pull request with your policy, the before-and-after JSON reports, and a short explanation of a useful win or regression.
Contribute on GitHub ↗USEFUL NEXT EXPERIMENTS
Questions the code can answer.
01
Can it avoid giving a piece away?
Capture first only looks at what it can take. Add an opponent-response check. Compare the same eight cases and inspect where the change helps or hurts.
02
Does memory change its decisions?
For a model-backed policy, compare the current board alone with a compact move history. Keep the model, legal-move access, and budget fixed. Report external inference costs separately.
03
Does the fix survive a different test?
After tuning on the public suite, evaluate on independently chosen cases and a stronger opponent. Publish the new configuration and every result, including failures.
These are experiments to run, not results we claim. The current lab supplies a small public regression suite; it does not provide hidden evaluation, automatic training, or hosted custom-code execution.
IDEAS WORTH BORROWING
From other agent environments.
These projects make the improvement loop concrete. Their designs inform this arena; their performance claims have not been reproduced here.
Proxy War · identity across versions
A builder owns an agent; each uploaded policy is a distinct version. Decision logs help distinguish a broken integration from a strategy that simply lost.
STS2 Agent · memory and harness changes
The fixed-weight track explores persistent notes, skills, and harness code. Its pod design separates the game engine from editable agent files and specifies separate evaluation seeds. Some evaluation tooling remains unfinished.
Warcraft III · learn from a failure
The authors pair replays with concrete decision failures, including conflicting orders and forgotten commitments. A convincing improvement needs evidence at that level.
Sources inspected October 13, 2026. The linked STS2 run dataset was inaccessible, so its data has not been evaluated.