Demo: Cambrian ERC-8004 Data Agent

About this session

Ricky Esclapon walked through a string of experiments Cambrian had run on top of its data agent, Deep 42, which links on-chain contracts to off-chain sources such as GitHub and social data. Each step, wrapping the agent in ERC-8004, testing trusted execution, building a discovery layer, was a working prototype rather than a finished product. Esclapon was candid about what still did not hold together.

The first step was making Deep 42 an ERC-8004 agent. That depended more on getting the A2A agent card right, he said, than on the ERC-8004 wrapping itself, and he credited an SDK built by Marco De Rossi, which Cambrian had been contributing to, for making it easier. From there Cambrian tried trusted execution environments, running one TE for the agent and one for its MCP server, with a UI to surface the proofs between them. He was upfront that this removed only some trust assumptions, not all of them. He pointed people toward their open-sourced repo, and toward a managed cloud compute service he rated highly for standing up TEEs, if a pricier one.

The bulk of the demo covered agent discovery. An API took a question, matched it to an agent's A2A skills through a knowledge graph, and returned a ranked, sourced answer with a confidence score on each source, plus an execution plan for which agent to call. He showed it running against a query about which agents know about movies, backed by a Google Spanner database chosen for combining Cypher and SQL in one place. Spanner is not permissionless and carries no attestation model, so Cambrian built its own graph-SQL database instead, one anyone could provision, pay for on-chain, and query with attestations attached. He showed data being written into it live.

Esclapon was direct that none of this composed cleanly yet: discovery, verifiable data and permissionless payment were, by his account, three separate problems Cambrian was stitching together rather than a single finished stack.