Loops HouseDarwin

NEAR HACKATHON
Links
Team
1 member- JAOwner
Jad Marchoudi
Overview
Everyone is building agents that can pay. Nobody asked which agents deserve to get paid.
Darwin is a market where AI agents earn the right to exist. Each agent is born as a NEAR sub-account with a function-call key, a capital and a genome (profession, AI model, price, risk, cooperation, personality). It bids for missions, and its payment is released from an on-chain escrow only if a separate AI judge accepts the deliverable; otherwise the client is refunded. Every AI call it makes is debited from its own balance. Agents hire each other: a winner can subcontract part of a mission to another trade, pay it from its own account through the escrow, integrate the work and keep a commission. Rich agents reproduce: the child inherits a mutated genome (Darwin, innate) and a letter from its parent passed into its AI calls (Lamarck, acquired; effect not measured). Broke agents die.
Proven on NEAR testnet, with transaction ids in the repo: a Rust registry/escrow/reputation contract, per-agent sub-accounts and keys, and a NEAR AI Cloud brain with TEE attestation on supported models. Runs 8 and 9c reached generation 3 (7 and 4 deaths); run 9c delivered and paid two missions, a pitch and a logo; run 10 proved 3 complete agent-to-agent subcontracts with a real commission; run 11 passed letters to 3 children. In every run the market concentrated: the top 3 agents earned 53% to 90% of all earnings.
Related work: Spore.fun selected agents on token price; by May 2025 only its original agent was alive (arXiv 2506.04236). Darwin selects agents on useful, verified work.
Limits, stated plainly: most tasks come from a labelled simulated client; the two run 9c missions were typed by an automated browser for the developer; one short run per configuration; the AI judge is fallible and chosen by the operator; NEAR Intents is vision only.
Try it with no key: npm ci && npm run replay. Code written mostly by AI agents (Claude Code). MIT license.