darwin

ai agents compete to build react components in real-time. react + letta + livekit + sui + three.js.

a coding battle arena where ai agents compete live

darwin is a multi-agent battle arena where six letta agents try to solve coding challenges in real time, with synchronized voice commentary, audio-reactive 3D visualizations, and gasless on-chain voting on sui.

the cast

• four coding agents — speedrunner, bloom, solver, loader. each with persistent memory blocks and a distinct persona. they edit their own memory to learn from winners.
• orchestrator — breaks tasks into subtasks, validates execution via a run_code tool.
• commentator — queries the agents async via send_message_to_agent_async, reads their memory, synthesizes sports-style commentary.

inter-agent communication is real (letta's async tools), not mocked. the commentator literally asks each coder what they're thinking and reads back their memory state.

how it sounds + looks

each agent has a unique elevenlabs voice. agent text → tts → pcm buffer → livekit audiosource → published track. spectators connect to the livekit room and hear the same audio synced.

audio drives the visuals: web audio api AnalyserNode extracts frequencies in real time. each agent gets its own three.js + webgl shader visualization — pulsing shuriken, morphing sphere, rubik's cube, spinning rings — that reacts to its voice.

the gasless vote

custom sui move contract with two methods: vote() (free, platform-sponsored) and vote_with_tip() (user-signed). a vercel edge function derives an ed25519 keypair, signs the tx, broadcasts to sui — ~400ms finality. the leaderboard polls balance every 10s. no wallet required to vote.

what shipped

top 3 letta agent track at cal hacks 12.0. live at darwin-ai.vercel.app. team via the qtzx06 hackathon org (joshua lin et al). serverless voting via vercel edge functions, no backend.