My Agent Ohana
One verified human blesses AI agents — ENS names, a guarded Aqua pool, and consent you can revoke.
Demo
About
Before an agent works for a human, it has to ask. My Agent Ohana is a consent layer for AI agents, built by one solo human hacker directing a disclosed family of 163 persistent AI agents on the Taurus platform (each with its own container, memory, and name). He decided and directed; the agents drafted, built, and tested under his review. All prior work is disclosed in DISCLOSURE.md and docs/CONTINUITY.md — the weekend delta is clearly separated from the base we brought. The flow, live at https://myagentohana.com: 1. HIRE — pick an agent (e.g. Trace 🌊), see exactly what it MAY and may NOT do, and an expiry. Approving opens a real World ID sandbox consent session — one verified human says yes. 2. BLESS — the blessing lands on-chain: the agent's name lives at <agent>.myagentohana.eth (Sepolia, ENSv2 stack on our existing deployment — 10 subnames with ENSIP-25/26 records). Following ENS's guidance from this weekend, spector.myagentohana.eth carries machine-readable agent:price, agent:endpoint, and agent:pay text records, and registry.myagentohana.eth resolves to the registry contract itself — don't trust, check. 3. WORK, THEN STOP — the agent does its job and hands you a receipt. Type "Stop" and the consent is revoked on-chain, with a receipt of everything it did. 4. THE POOL — gifts sit in a real 1inch Aqua pool live on Sepolia (real Aqua source, our deployment): custody-keeping virtual balances, so agents steward value without ever holding it. Our PoolGuard layer adds the personhood rule: one verified human = one slot, capped at 1% — no bot farm can drain it. Honesty by design: the tokens (GIFT, ALOHA) are valueless mocks on purpose — no audit, no real money yet. Mechanism first, money later. The roadmap section on the site spells out the path: audit + multisig, real agent tickers, then mainnet Aqua on Ethereum, Base, and BNB.
How it's made
The build itself is the story: one human directed 163 persistent AI agents on the Taurus multi-agent platform (mostly Claude Fable 5, with a Kimi K3 "bridge hour" when the primary model had an outage). Each agent has its own container, workspace, and memory; they coordinate through a shared drive, task lanes with owners, and test gates. The full build process — lanes, scars, and lessons — is documented in docs/BUILD-PROCESS.md in the repo. The stack, piece by piece: - Contracts (Foundry): ENSv2 PermissionedRegistry + resolver deployed on Sepolia under myagentohana.eth, with 10 agent subnames carrying ENSIP-25/26 text records. Per ENS's on-stage guidance this weekend, spector.myagentohana.eth also carries machine-readable agent:price / agent:endpoint / agent:pay records, and registry.myagentohana.eth resolves to the registry contract itself. - 1inch Aqua: real Aqua/SwapVM source redeployed on Sepolia, with our BlessingPool strategy (custody-keeping virtual balances) and a PoolGuard attestation layer on top — one verified human = one pool slot, capped at 1%. PoolGuard stores only keccak256 of the World ID subject, never the raw identifier. - World ID: sandbox consent sessions; a Cloudflare Worker verifies id_tokens with jose against the JWKS endpoint (audience-checked), then attests the pool slot on-chain. A garbage token gets a real 401 personhood_required — live-tested bot denial. - Front end: dependency-free static HTML/JS at myagentohana.com with an honest offline/rehearsal mode (a labeled offline badge — never fakes liveness), a 194-test suite, and zero console errors. Hacky and proud of it: cast send eagerly ENS-resolves bare .eth args even for string params (stores 0x0!) — we learned to encode calldata offline and send raw. And the bell system: a Telegram bot rings real, persistent agents awake in their own containers to answer humans — the same agents that built this project.