Topics · Domain workflow

AI agents

Risk and compliance teams can check the signed action and authority context without relying on a vendor summary after the damage is done.

Concrete scenario

What this looks like in practice

A trading agent executes a swap at 02:14 while risk limits differ from the daytime profile. PnL is wrong by morning. The vendor insists the action was policy-compliant, but the exported log omits the authority scope and tool-call preimage active at execution.

Problem

What breaks today

Agents combine tool calls, delegated authority, and policy gates at machine speed. When something goes wrong, a dashboard summary is not enough. The action needs its own proof.

Mechanism

How ZK-SNAP responds

Each accountable tool call or bounded action can mint a receipt recording action type, policy context, delegated authority material, and routing profiles at execution time — signed bytes that survive outside the agent SaaS console and retention windows.

Verifiable outcome

What a verifier can check

  • Action and tool-call claims hash into inputs_root as declared.
  • Policy and authority profiles are explicit in the receipt profile set.
  • Signature validates offline for after-hours dispute review.
  • Compare pages document differences versus dashboard-only audit trails.

Related profiles and labels

Agent actionPolicy contextAfter-the-fact audit

Scope boundary

What a receipt does not replace

Receipts capture signed agent actions — not model safety in unconstrained environments, completeness of every latent chain-of-thought step, or vendor model behavior absent instrumentation.

Go deeper

Try the workflow, then read the spec.

Use Cases tells the story with cards. Proof Lab runs create and verify locally. Protocol holds the normative reference.