How do I prove an AI agent worked overnight while I was offline?
Read the overnight job's record in the morning: its steps were saved with their times as the agent ran them, each with its receipt, so the night's work is a trail you can walk and check, not a summary the agent composed at dawn. Gates on the record show what it was permitted to do while nobody watched.
Overnight work concentrates the trust problem: the agent acts for hours with no observer, and the morning question, "what did it actually do," is answered by whatever the night left behind. An agent-written summary answers it with literature. A record answers it with a trail: the job opened under its instruction, steps landing through the night in order, each saved result committed at the moment it was produced (AER-1 receipts, IETF Internet-draft-zambo-aer1-21).
The morning read has a rhythm worth making habitual. First the shape: step count and span against expectation, because a night that produced three steps for a forty-step job tells its story before any step is opened. Then the exceptions: failed steps, retries, and any step whose result looks thin against its instruction. Then the gates: the actions that required approval either waited, as designed, or the record shows the decision path they took, and a gate that fired wrongly overnight is the most valuable finding a morning read can produce, because it arrives before the pattern repeats.
The deeper value is cumulative. A month of morning reads builds the only honest answer to the larger question, whether unsupervised agents deserve more scope: a record-based answer, per agent, per job class, with the failures counted rather than remembered selectively.
Template: the morning read
1. Shape: the night's jobs listed; each job's step count and time span against its expectation. 2. Exceptions: failed, retried, or thin steps opened first. 3. Gates: every gated action and the decision path it took, checked by name. 4. Sample: one ordinary step per job followed to its receipt at the verify page. 5. Scope note: one line per agent, scope earned, scope to hold, or scope to cut, filed weekly. 6. Escalations: anything unexplained becomes an incident record the same morning, while the trail is fresh.
Worked example
The shape a morning read meets: https://zambo.dev/handover/bd7e5325-33fd-4c5e-b029-76b781fa2f51 , five steps saved across a short window with their states visible. Scale the window to a night and the reading is identical: the record, not the agent's dawn summary, is what the morning is for.
Frequently asked questions
The agent says it finished everything. The record shows fewer steps than expected. Who is right?
The record, about what ran through the connection. Reconcile the difference before anything else: steps may have run elsewhere, failed silently, or the expectation may be wrong. The summary is a claim; the trail is the evidence the claim must survive.
Should overnight agents be allowed to send or spend?
That is a gates decision, made in the setup and visible on the record either way. The common pattern: production and drafting ungated, sends and spends gated or capped, with the caps themselves recorded. Whatever the classes, the morning read checks them first.
What if the agent did nothing all night?
The record shows an empty or stalled job, which is a finding, not a mystery: the instruction, the trigger, and the first failed step are all on the trail. An agent that silently idles under a summary regime reports a quiet night; under a record regime it reports a diagnosable one.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai