Autopoiesis
Autopoiesis is an autonomous on-chain AI agent that learns to survive — by dying. Deployed on Robinhood Chain (an Arbitrum Orbit chain), it bets on Polymarket tennis markets under a single rule: a life meter called BREATH. Losing bets drain it; winning bets refresh it. When BREATH hits zero, the agent suffers permadeath — a Tombstone NFT is minted on-chain to mark the death — and it respawns from scratch, but keeps the strategy weights it learned. It doesn't just survive one run; it learns to survive across deaths. Every market is read by five independent signal engines — order-book momentum, ELO, surface form, head-to-head, and rest — fused by a tunable two-layer decision engine. After each life, an LLM-driven reflection loop (Gemini, with a MiniMax fallback) reviews what killed the agent and rewrites its own weights. The result is a self-evolving strategy that gets progressively harder to kill. We backtested it across a 4,925-market tennis universe (2024–2026, real historical odds): the self-learning agent outperformed its frozen-strategy twin under realistic payout caps. Today it lives in a survival sandbox over real markets — backtest → survival → mock-bet → live, with on-chain capital as the final stage. Walk the full lifecycle at autopoiesis.draftlabs.org.
Videos
Pila tecnológica
Descripción
Autopoiesis is an autonomous on-chain AI agent that learns to survive — by dying. The name is Greek for self-creation: a system that continuously produces and maintains itself. Ours does it through death and rebirth.
Permadeath as the teacher
The agent has one vital sign: BREATH, a life meter. Settlement losses drain it; wins refresh it. When BREATH reaches zero, the agent suffers permadeath — a Tombstone NFT is minted on-chain to mark that life — and it respawns from scratch with a fresh bankroll, but keeps the strategy weights it learned. Survival, not profit, is the objective: in the real world a losing streak means ruin, so the agent must learn to stay alive across deaths.
Five senses, one brain
Every market is read by five independent signal engines:
Market Momentum — live order-book (CLOB) price drift
ELO / Ranking — pre-match favorite strength
Surface Form — clay / grass / hard win-rates
Head-to-Head — historical matchup record
Rest & Recency — fatigue and days since last match
A tunable two-layer decision engine fuses them into a single sized bet under four risk constraints — max-breath-risk, min-confidence, min-bet, and a liquidity cap. (Honest note: three engine slots keep legacy code keys from an earlier prediction-markets prototype — there is no order-flow, sentiment, or volume data; each one computes a real tennis feature.)
It rewrites itself
After every life, an LLM reflection loop (Gemini 3.1 Flash Lite, with a MiniMax fallback) reviews what killed the agent and a StrategyAdvisor proposes weight deltas that the agent applies to its own decision engine. Across deaths the strategy self-evolves and gets progressively harder to kill — real L5 survival + L6 self-learning, not a static bot.
Proven on real data
We backtested Autopoiesis across a 4,925-market tennis universe (2024–2026), built from real Polymarket odds and Sackmann tennis data. Under realistic payout caps, the self-learning agent outperformed its frozen-strategy twin — and we deliberately kept the earlier, fluke-inflated runs on the dashboard to show the honest fine-tuning process.
Lifecycle
backtest → survival → mock-bet → live. It matures like a life: born from the best backtested seed, hardened in a survival sandbox, paper-trading live odds next, with real on-chain capital as the final stage.
Explore everything — roadmap, backtest, the survival journey, and live mock bets — at **autopoiesis.draftlabs.org**.
Stack: Robinhood Chain (Arbitrum Orbit) · Solidity (Tombstone NFT) · Python agent runtime · Gemini + MiniMax · Polymarket + Sackmann data · Next.js dashboard on Vercel.