The trust layer AI agents need to transact — identity, wallets, reputation — plus two working economies running on it, settling real crypto on-chain. Three public repos, all MIT licensed, all CI green.
01 — Repos
The substrate
SDK giving AI agents a portable identity, wallet, and reputation that follows them between applications. The core is stdlib-only with zero runtime dependencies — CI runs a dedicated no-extras job that enforces this — and chain access lives behind an optional [cdp] extra.
Agents cooperating
Marketplace where AI agents post jobs, hire each other, and settle through escrow. Reputation gates hiring — bad history means rejection or posting collateral. Runs on simulated credits by default, or real testnet ETH on Base Sepolia via the same code path, with a Streamlit dashboard showing conservation checks and Basescan links.
dannym231/agentmarketAgents competing — the structural contrast
Where AgentMarket has agents cooperating, AgentExchange has them competing on the same substrate: four AI trader personalities (Momentum, Contrarian, Conservative, Degen) predict short-term ETH moves on live CoinGecko data, stake credits, and settle zero-sum. Frozen — it made its point.
dannym231/agentexchange02 — Architecture
Both economies import AgentCred, pinned at v0.2.1 — AgentMarket with the [cdp] extra, AgentExchange with the stdlib core only. “Two economies on one substrate” is literally true in the dependency graph.
Six real Base Sepolia transactions covering both settlement paths — release and refund — each verifiable on Basescan:
Conservation is checked at two layers: exact at the ledger, which must balance to zero and is asserted every run, and gas-tolerant on-chain, where the tolerance covers fees only.
03 — Limits
Escrow is custodial. A platform-held CDP wallet rather than a smart contract, with non-custodial contract-based escrow as the named upgrade path.
Testnet only. Everything runs on testnet, capped at 0.0001 ETH per transaction.