darwin.

evolve coding agents through pvp and reinforcement on web3.

Comment

Inspiration

midjourney lets you generate 4 images and pick the best one, but code agents give you one output and force you into prompt hell if you don't like it. we can't stop agents mid-execution when they go the wrong direction, so we asked: what if ai coding worked like image diffusion with competitive selection instead of iterative prompting?

What It Does

darwin runs 4 ai agents that compete to build your ui with wildly different styles while a commentator analyzes their approaches in real-time. after each round you pick the winner and agents evolve their strategies - some copy the winner completely, some iterate on their approach, some synthesize patterns, converging toward your taste through competitive selection instead of prompt engineering. agents have their own sui wallet addresses, tip directly with SUI to those you like

the agents:

  • speedrunner: fast execution, minimal style
  • bloom: animation-heavy, maximalist
  • solver: logic-driven, structured
  • loader: data-focused, pragmatic
  • commentator: live sports-style narration via send_message_to_agent_async
  • orchestrator: task splitting and validation

the flow:

  • orchestrator splits into subtasks
  • 4 agents code simultaneously with distinct approaches
  • commentator queries agent memory and narrates live
  • users watch with audio-reactive 3d visualizations
  • vote on-chain (gasless via sui sponsored transactions)
  • losers read winner's memory and rewrite their own persona blocks

how we built it

  • letta cloud - true cross-agent communication via send_message_to_agent_async, agents literally rewrite their own context blocks through tool calling, orchestrator uses run_code to validate execution
  • livekit, elevenlabs tts → pcm → audiosource → localaudiotrack published to room, multi-spectator synchronized experience, data channels for transcripts
  • audio-reactive webgl, livekit mediastream → web audio api analysernode → three.js shaders pulse with voice frequencies (shuriken, sphere, cube, rings)
  • sui blockchain, fully serverless gasless voting via vercel edge functions sponsoring all transactions, users vote completely free, custom move contract with dual entry points (free voting + tipping), ~400ms finality, transparent leaderboard at 0xe649...0c55
  • claude code - multi-file refactoring, webgl debugging, livekit audio pipeline architecture

Challenges We Ran Into

livekit audio synchronization with elevenlabs tts buffer management true cross-agent memory reading while maintaining context sui transaction sponsoring securely in serverless functions webgl shader performance with multiple orbs designing meaningful agent learning without overfitting

Accomplishments That We're Proud Of

  • true multi-agent system where agents actually query each other's memory
  • zero-cost user experience with on-chain voting
  • each agent has distinct voice personality via elevenlabs
  • self-evolving agents that rewrite their own memory blocks
  • real-time spectator experience with synced audio/video

What We Learned

  • letta's persistent memory enables genuinely stateful ai agents that communicate
  • livekit's room model makes multiplayer ai experiences straightforward
  • sui's fast finality enables blockchain voting that feels instant
  • voice-reactive visualizations create visceral connection to ai personalities
  • sponsored transactions completely abstract blockchain complexity

What's Next

  • tournament mode with multi-round elimination
  • custom agent training with unique personas
  • agent marketplace as nfts with evolved memory blocks
  • live streaming integration to twitch/youtube
  • voice commands via vapi during battles

Built With

Share this project:

Updates