Episode 233: Agents have keys. Who checks what they did with them?

Identity work has gotten good at saying who an agent is: keys, delegation, mandates, verifiable credentials. It says much less about what the agent then did, and that's where disputes, audits and payments actually happen. Steven Mih walks through an open approach: each party to an agent action seals its own signed record, the two records cite each other, a public log fixes when they existed, and anyone can verify the result offline with no account and no trust in the operator. We'll cover how it builds on SCITT (RFC 9943) and COSE receipts, how it binds to identity you already have (did:web, owner-signed agent keys, AP2 mandates), a live example on Mesh-LLM, and the hard questions: what a record proves, what it doesn't, and why it should report checkable properties instead of a trust score.

#Speakers

• Host: Mike Schwartz, Founder/CEO Gluu
• Guest: Steven Mih, Founder & CEO Action State Group

#Homework

• The open format, plus the trust map of what each check shows and doesn't
• The Internet-Draft (-05)
• SCITT, the transparency layer it builds on

#Takeaways

• An agent’s diary is not evidence. Steven’s invoice-processing agents exposed the problem: CFOs needed to stand behind the work, but ordinary logs could be missing or changed. His approach captures signed, tamper-evident records as actions happen.
• A signed lie is still a lie. Cryptography can protect a statement from alteration. It cannot prove the statement is true. If an agent claims it paid an invoice, ask for evidence from the receiving side.
• Capture the history before you know the questions. Record who delegated the work, what was authorized, what checks occurred, and what the agent did. Preserve that context from the start so auditors and investigators have something to examine later.
• Agent oversight needs a regular review cycle. Steven recommends daily evaluation of recorded actions, with weekly expert review of a sample. For judgments such as whether an agent pressured a customer, use an explicit rubric and check whether automated evaluations agree with human experts.
• Prove the outcome—and how the agent achieved it. Logins, usage, and user satisfaction give an incomplete picture of agent performance. Steven’s business case is evidence that the job was completed under agreed criteria. That supports measuring value and paying for outcomes.