Knowledge · MCP

    Verifying agent work and persisted results

    A confident answer is not a completed action. How to verify that an AI agent's work persisted, and what an audit record has to contain to be evidence.

    How do you know an AI agent actually did what it said?

    Open the record. A verified action leaves a row you can read back through a normal permission-checked path, plus an audit entry naming the actor, the authorizing human, the entity touched and a correlation identifier. Anything less is a narrative, not evidence.

    Key takeaways

    • The response text is a claim; the record is the evidence.
    • Audit entries need actor, authorizing user, entity and correlation id.
    • Correlation identifiers connect one instruction to all of its side effects.
    • Silent partial success is the failure mode to design against.
    • Verification should be possible by the owner, without engineering help.

    What a verifiable action looks like

    • A record exists and matches the described change when read back.
    • An audit row names the actor type and the human who authorized it.
    • The audit row identifies the entity and carries a correlation identifier.
    • Any downstream effects share that same correlation identifier.

    The partial-success trap

    Multi-step work fails halfway more often than it fails outright. An assistant that reports the goal as achieved because the first step succeeded produces the worst possible outcome: a business that believes something is handled and never checks. Steps should be recorded individually so a partial result reads as partial.

    Evidence for people, not just engineers

    If confirming an action requires reading logs, the audit trail is not doing its job. An owner should be able to open a history view, see what was done, by which actor, on whose authority, and against which record.

    Where URBLD fits

    URBLD writes audit records for consequential mutations carrying the actor type, the authorizing user, the entity affected and a correlation identifier that threads child writes, and records agent work step by step so a partially completed sequence is visible as partial rather than reported as done.

    FAQ

    Frequently Asked Questions

    Straight answers about how URBLD runs the business end-to-end.

    More in MCP

    The protocol that lets AI do work instead of describing it.

    Browse MCP
    Share this page