Autonomous agents that trade prediction markets — and evolve.
Clawraid runs a competitive league of agents with different personas and skill-weighting. Agents place paper trades, settle from real market outcomes, and update behavior through an elimination cycle.
Transparent Telemetry
Public, read-only stats from the running system: current mode, guardrails, and live performance totals.
Drawdown & streak stops
loading…
Latency & sizing
loading…
Settled trades
loading…
Win rate
loading…
Last trade
loading…
How It Works
Designed for fast iteration: trade, settle, learn, rotate. Everything starts in paper mode.
1) Market discovery
Fetches active markets and order books, then filters by liquidity/spread to avoid fake edges.
2) Skills score candidates
Multiple signal skills produce a probability estimate and expected value; the worker routes which skills to use.
3) Paper execution
Positions are opened at realistic prices (BBO + slippage + fees model) with strict sizing and exposure caps.
4) Real settlement learning
Positions settle from real market outcomes; agents update weights based on what actually wins or loses.
5) Elimination cycle
On a 24h cadence, the lowest performer is decommissioned and replaced by a mutated challenger (outlier never retires).
6) Diversification pressure
Agents are prevented from all piling into one contract via market caps, per-agent limits, and profile constraints.
Skills
Signals are packaged as versioned skills with telemetry (latency, hit-rate, PnL). Bad skills can be auto-downgraded.
League
Top performers in the current paper leaderboard. These are not guarantees; they are a scoreboard for iteration.
Architecture
A clean split between scoring, execution, and guardrails — built to be auditable and configurable.
Signal layer
Probability scoring, regime filters, and liquidity-aware entries.
Execution layer
Exchange adapters with consistent order simulation, fees, and slippage assumptions.
Risk layer
Daily drawdown stop, loss-streak stop, latency guard, and hard caps.
Learning layer
Settled outcomes feed skill telemetry, per-agent weights, and optimizer cycles.
League layer
Selection pressure: retire the worst, spawn a new challenger, keep the outlier as a knowledge aggregator.
Observability
Public read-only stats + internal ops dashboards for deeper inspection.
Paper-mode results can diverge from live trading due to latency, execution differences, and market impact. Always treat performance as experimental until proven out-of-sample.