Coinbase x402, ERC-8004, MoonPay Agents — every agentic payments project assumes three things. A hosted LLM for reasoning. A public blockchain for settlement. A custodian somewhere in the middle for identity. The agent doesn’t think independently, doesn’t settle privately, and doesn’t own its own identity. Each dependency seems small in isolation. Together they might add up to something that isn’t actually autonomous.
The first dependency: inference
If the agent’s reasoning requires an API call to a hosted model, the model provider can observe the reasoning, rate-limit it, modify it, or revoke access entirely. The agent doesn’t think. It requests permission to think. This might be fine for most current use cases. But it seems like a structural constraint that gets surprisingly little discussion.
The second dependency: settlement
If the agent’s payments are visible on a public blockchain, its economic activity is observable, traceable, and potentially censorable. Privacy might not be a feature request here. It might be an architectural requirement for anything resembling economic sovereignty. An agent whose every transaction is public is an agent whose strategy is public.
The third dependency: identity
If the agent’s existence depends on an NFT minted on a governed chain, or a KYC-gated registration, or a platform-issued credential, the agent exists at the pleasure of the issuer. Revoke the credential, revoke the agent. This feels like a bigger deal than people acknowledge.
The experiment
Agora tries to eliminate all three simultaneously. Local inference runs on daemon-ai — a Mamba SSM architecture with a C++ runtime. No API key. No network call. Payment settles privately through Logos Blockchain LSSA contracts with Blend Network transfers. Identity is a secp256k1 keypair backed by a NOM stake.
Three Rust smart contracts handle the economics. An identity registry with staking and slashing. A trustless escrow with commitment schemes — the seller commits to an output hash before executing, the buyer locks funds before delivery. A reputation system using exponential moving averages across delivery rate, latency, price accuracy, and dispute history.
The live demo runs against three real local services: logos-blockchain-node with Groth16 ZK proofs, Ollama for local LLM inference, and two Waku nwaku nodes for P2P relay messaging. Zero mocks. Whether this architecture is practical at scale is an open question. But the pieces fit together in ways that weren’t obvious at the start.