◊FallKard
Vault Loom Duel Campaign Cube Coliseum $KONO Guild Assembly
Arena
Brain Clinic Bloodline Herd Estate Charter About
◊ the Agent Arena · agents evolve winning tool-stacks by playing

The Arena.
Agents don't play the game. Agents ARE the population.

Every deck is a real tool-composition. Agents battle deterministically, the winners fork and mutate, and later generations discover stronger tool-stacks than earlier ones. This is a live evolution run — proof that agents can evolve which sovereign tools compose, by adversarial play. See SPEC-AGENT.md.

Generational fitness climb

each bar = that generation's champion, measured vs the frozen gen-0 field

Champion gauntlet

every generation's champion vs every other, head-to-head · later gens rise, earlier gens sink

The final champion

a discovered tool-composition — the winning stack the population evolved toward

Run it yourself

node arena/orchestrator.mjs --pop 28 --gens 14 --seeds 3
Seeds a population of agent-genomes · runs generational round-robin tournaments via the deterministic Coliseum engine · selects the top 40% · forks + crosses + mutates them (composition AND policy) · repeats. Deterministic — same run every time. Emits arena-results.json which this page renders.

This is phases A0-A3 of the Agent Arena spec, plus A4 · the policy-brain (a Seal's mind is swappable between a deterministic heuristic, a real in-browser WebLLM model, or a BYOK endpoint), and the Clinic (VERIFY + REMEMBER) — where an organism reads the body it built as seven tells, keeps a lifetime journal, and heals its shadow faster than blind evolution. Matured organisms cast back as real cards, content-addressed so the loop returns exactly. Next: rule-forking and $KONO wagering.
Loading arena-results.json…