Nobody Verifies What Your Agent Reads
2026-11-03 –, Lightning Stage

AI agents are now one of the largest new consumers of Ethereum data, and not one of their reads is verified. Every blockchain MCP server, mine included, is an unattested oracle feeding financial decisions. You'll leave knowing where the verification stack and the agent stack failed to meet, and what closing that gap actually requires.


We spent years building verification for Ethereum: light clients, storage proofs, TEE attestation. In parallel we built an agent stack that verifies who an agent is (ERC-8004 identity) and attests what it did. But, nothing verifies what it read.

I run indexing infrastructure across several networks. I'll share what that seat shows: how agents actually query chain data, where what they read silently diverges from consensus (reorgs, unfinalized state, transformed metrics nobody can re-derive), and why recent benchmarks find these errors undetectable downstream. Models stay fluent while being wrong.

Envio is in the critique. A handful of hosted indexers and RPC providers now decide what agents believe about Ethereum, and open source is not verification.

I'll close with a concrete proposal for attested reads and what MCP would need to carry provenance.

Agent steward.
Loves being a dev.
Envio co-founder.
Safe Guardian.
Anarchist at heart.
Been building since gas prices allowed experimentation on mainnet 😅.