An AI agent isn't a single call. It's a sequence of choices: read this, route to that model, call this tool, remember that fact, spend here, hand off there. Each choice changes what happens next. When something goes wrong, or when someone asks why the agent did what it did, the honest answer is usually a story pieced back together from logs the operator controls. That's a claim, not proof.
AI agent receipts close that gap by signing off on the agent's actions as they happen.
What an agent receipt captures
An agent receipt locks in the conditions of a single agent action into a signed record. Across a run, those records form a chain: a timeline you can check for what the agent actually did.
- Route decisions: which path or model the agent picked at a branch point, and against what inputs.
- Tool calls: which tool got called, with what arguments, at what step.
- Model choices: which model handled which part, so a downgrade or swap can't happen quietly.
- Memory and state: what the agent carried forward, so the history survives across steps.
Each one is signed with post-quantum ML-DSA-65 and anyone with the public key can check it offline. The receipt confirms what happened. It doesn't assess whether the decision was the best one. That's a narrow, honest claim, and it's exactly what makes it hold up later.
Why a chain, not a single stamp
Agents get assessed on the whole trip, not one hop. A single signature over a final output tells you nothing about the seven decisions that produced it. A receipt chain lets a reviewer walk through the sequence, decision by decision, and confirm none of it got changed or added after the fact. Think of it this way: providers own the roads the agent travels on, the agent owns the trip, and the receipt is the record of the route.
See it run
The clearest way to understand agent receipts is to watch one get built. Agent Trip streams a live sequence of route decisions, minting and checking a receipt at each step. It's the agent-action side of the same receipt system that signs inference and human-origin data. For one concrete example, the Leanstral receipt example shows a SmartAgent receipting one live Mistral model-call route: provider, model, prompt and response hashes, timing, and a signature that checks out. The Model Receipts hub covers how the same receipt works across different model providers.
How much proof does the task need?
Not every agent action needs the same amount of proof behind it. A low-stakes lookup and a regulated financial handoff sit at very different points on that scale. Hive's AI assurance tiers use the same receipt system across levels, from signed inference to verified agents to regulated workflows, so you can choose how much independent, outside proof a given task actually needs. Checking a receipt is always free. You only pay to sign one.
Watch route decisions get signed and verified in real time, then see how assurance tiers scale the proof to the task.