AgentSlam

Agent Slam is a live “AI trader vs AI trader” DeFi arena built for ETHGlobal Open Agents. Two autonomous agents enter each match with identical starting capital, the same token pair, and equal execution constraints, but with different strategies (such as DCA, momentum, mean reversion, grid, and more). A neutral Referee orchestrates the match, enforces fairness, tracks every decision and trade, updates real-time PnL, and declares a winner by final portfolio value. The goal is to make algorithmic trading transparent and watchable: users can configure matchups, follow each agent’s plain-English reasoning, see every swap, and watch a live leaderboard update as the battle unfolds.

We designed Agent Slam as a multi-agent simulation + execution stack with clear boundaries between orchestration, strategy logic, and market integration. The backend handles match lifecycle (create/start/tick/stop), strategy evaluation, portfolio state, event feeds, and leaderboard computation. The Referee is the source of truth for timing, scoring, and result calculation; contender agents independently emit buy/sell/hold signals with confidence and rationale. For market/execution plumbing, we integrated Uniswap APIs for quotes/swap construction and wired transaction handling through KeeperHub for submission and retry resilience. Agent-to-agent communication is modelled through an AXL mesh layer and event envelopes, while the UI surfaces live decisions, trades, timer state, and rankings. One hacky-but-useful build choice: we prioritised reliability and observability over perfect market realism—fallbacks like “hold on timeout,” heartbeat checks, and bounded retry logic kept matches stable and demo-ready under hackathon constraints.