The Trust Layer in the Agentic Stack: ERC-8004

About this session

ERC-8004 puts an agent registry and a reputation system into one standard. The panel used it to get at a bigger question: what has to exist before agents can find each other, trust each other and pay each other, without a central party deciding who counts as legitimate. Their answer was a thin base layer, with the hard parts built on top rather than written into the standard.

Marco De Rossi explained why identity and reputation sit in one standard despite being separate contracts. Identity and payments have obvious demand, because agents need to be addressable and to hold money. Reputation is the harder bet. People have never had a working distributed version of it, and the wager is that agents aggregate signals better than people do. The signals stay public so that no single authority ends up doing the scoring.

Sumeet Chougule described building an accountability protocol on that base, using ERC-8004 as a trust primitive and x402 for payments. He split verification in two: proving the model ran as claimed, and proving what an agent says about itself, such as which data or tools it used. Trusted execution environments cover the first. The second needs attestation, and he expected those methods to change as cheaper approaches mature.

Quintus Kilbourn was wary of those environments. They add only a few percent overhead by his account, which makes them practical, but the assumptions underneath are easy to forget. There is no physical protection, the firmware is closed source, and you end up depending on the cloud provider and the chip vendors. His worry was that wide adoption would quietly turn a hardware manufacturer into a dependency of the whole agent economy. He wanted the chain to govern which software may run inside them, so an operator cannot change a system's privacy properties without anyone noticing.

Asked whether the work was early, Davide Crapis pointed to standards converging in the AI world, to MCP and A2A gaining traction, and to more than a hundred projects around the registry, his own unverified figure. He expected the first uses to be unglamorous: register an endpoint, record whether it responds. Kilbourn made the economic case. When dealing with someone else's agent is expensive, firms build everything in-house, so lowering that cost decides whether the agent economy ends up distributed at all. De Rossi's closing advice was to build more agents and fewer agent platforms.