Execution receipt
A structured record an agent produces for each operation: what was read, which tools ran with what arguments, what was decided and what was written. Receipts make AI work reviewable and attributable after the fact.
Turning trust into checking
Without a record, delegating to an agent means trusting a black box. With one, the question changes from "do we believe the model?" to "does the record hold up?" — which is an ordinary reviewable problem. That shift is what lets teams widen an agent’s scope without simply becoming more optimistic.
What a receipt has to contain
What was read and from where, which tools ran with which arguments, what was concluded and on what basis, what was written or sent, and which approvals were granted by whom. Claims link to sources; actions link to approvals. Anything less is a log — useful for debugging, insufficient for accountability.
Receipts are for the accountable human
This is what separates them from logging. Logs are written for engineers debugging a system; receipts are written for the person answerable for the outcome. Different audience, different structure — and a system that only produces logs has not solved the verification problem no matter how thorough they are.
In AGNT
Receipts are AGNT’s core primitive — every run explains itself, which is what makes delegation safe.
Execution receipt — common questions
Is a receipt just a log with a nicer name?
No. A log records that events occurred; a receipt links claims to sources and actions to approvals, so a non-engineer can review the work.
Does producing receipts slow agents down?
Recording is cheap and happens as the run proceeds. The only deliberate slowness is an approval gate, which is doing its job.
How long are receipts kept?
They are yours and stored locally with the rest of the runtime, so retention is your decision rather than a vendor’s.