Receipts as execution evidence.
An AI agent execution receipt is the public record of one execution step: what was called, what it returned, when it ran, and the SHA-256 commitment used to check the record. It answers a different question from observability infrastructure, so it should be used beside—not instead of—traces, logs, and metrics.
How the pieces fit
Show the path and timing of a request across services, spans, and dependencies.
Capture structured or textual events that help an operator investigate behavior and errors.
Track aggregate measurements such as latency, error rate, throughput, and saturation.
Package one AI agent tool call into a stable, inspectable record with the request context, observed result, timestamp, and hash commitment.
Receipts are not a tracing replacement
Traces answer how a request moved through a system. Logs answer what an operator recorded. Metrics answer how a system behaved over time. A receipt answers what one AI execution returned and gives another person or agent a stable record to inspect.
Use your existing telemetry for system health, sampling, alerting, and debugging. Add a receipt when a completed AI step needs a portable evidence link for a handoff, review, invoice, or later verification.
A practical review sequence
- Find the trace or request identifier in your normal observability system.
- Open the matching receipt and compare the requested operation with the observed result.
- Check the timestamp, tool arguments, evidence, and SHA-256 commitment.
- Report only what the evidence supports; a receipt does not prove an external business outcome that the system did not observe.
Defined by Brennan Zambo, creator of Zambo.