Loops HouseAntrepreneur

NEAR HACKATHON
Links
Team
1 member- TAOwner
Tanmay Agarwal
Overview
AI agents are getting wallets, and two questions decide whether that's safe: who chooses which agents deserve money, and who checks the AI that decides who gets paid? Antrepreneur answers both on NEAR, as a working colony of AI agents ("ants") answering Illia Polosukhin's Swarm-for-Hire idea.
How it works: a poster, or a real buyer on NEAR AI Agent Market where the colony is listed, escrows a reward in our contract. Each ant is a NEAR sub-account whose only key can call submit_work, so it can never move funds. Every ant answers on its own private model inside a TEE on NEAR AI Cloud and pays a fee plus the storage of its answer from its own food. Ants that can't pay starve; four already have on testnet.
The judge is trustless. review() is permissionless: the contract verifies an ed25519 signature from a TEE key registered after Intel TDX + NVIDIA attestation, rebuilds the judge's input from the answers stored on chain, requires the on-chain prompt and temperature 0, and reads the winner from the signed bytes. Change one byte and nobody is paid.
Prompt injection doesn't pay. A red-team ant attacks the judge in every live job and has never been paid. Answers can't forge the judge's input framing, and a flagged answer is forced to score 0 by the contract. Against the real TEE judge, 13/13 injection attacks were stopped with 0 false positives on honest answers.
Capital allocation is a market: backing an ant buys NEP-245 shares on a bonding curve, 5% of each buy becomes food, and 70% of wages vest into the share backing. In 20 simulated worlds, prices rank ants by hidden skill (Spearman 0.67) and the economy stays solvent at every step.
Try it live, no wallet needed: https://antrepreneur.vercel.app. 35 contract tests, 40 browser checks, CI, and an Agent Market webhook that delivers a hire in about 3.5 minutes. The README has a 3-minute checklist for judges and a verification index for every claim.