What a brief is optimized for, and what it is not
A brief is optimized to be read in ninety seconds. It compresses a night of tool calls into the three things you might act on, phrases them confidently, and leaves the rest on the floor. That compression is the product. It is also where the truth leaks: the retry that finally worked reads as "done"; the source it could not reach reads as silence; the draft it wrote from a wrong assumption reads exactly like a draft written from a right one. None of this requires the agent to deceive. Summaries drift the way all summaries drift, toward the coherent version. The record does not drift. A receipt's hash was fixed at the moment of the call, and it recomputes or it does not.
The record exists; make the brief answer to it
When overnight work runs through the Zambo endpoint, the night's calls are all in one stream, each with its receipt. Counted on the evening of 2026-10-22, that stream took 500 receipts from 127 distinct callers in about five hours: prompt checks, capability searches, continuity reads, workflow closes. A brief built over a record like that can be audited line by line: this claim, this receipt; this claim, no receipt, and therefore a plan or a hope, not a fact. The ask to make of every brief, from any agent, in any app, is simple: show me the receipts under the claims. Agents whose apps cannot answer it are asking you to run your mornings on trust, and trust is the one thing the brief was supposed to reduce.
The two-minute morning habit
Read the brief for orientation, as intended. Then pick the claim with consequences, the payment, the sent message, the booking, and recompute its receipt at zambo.dev/verify. A match means the saved result has not changed since the work ran. Note any claim with no receipt beneath it and treat it as unverified until it has one. That is the whole practice, and it changes the relationship: the agent's summaries get more careful when the record is routinely consulted, the same way anyone's do. The brief stays pleasant. It just stops being the only version of the night.
The standing honesty note: a receipt proves the saved result was not changed after it was saved, never that the agent's overnight judgment was good. You still skim the drafts. You just no longer have to wonder whether the night described is the night that happened.
Questions, answered
What did my AI agent actually do overnight?
The brief tells you what the agent says it did. The record tells you what it did. If the overnight work ran through a receipt layer, every call is in the record with its time and the hash of its saved result: in one counted evening window on 2026-10-22, the Zambo endpoint recorded 500 receipts from 127 distinct callers in about five hours, the raw material any honest brief is built from. Ask for the receipts under the brief's claims, recompute one at zambo.dev/verify, and the difference between a faithful brief and a flattering one stops being a matter of faith.
Why is a morning brief not proof of the work?
Because the brief is written by the worker. It selects, compresses, and phrases, which is what makes it readable, and every one of those acts can drift from the record without anyone lying: a failed step becomes 'handled', a skipped check disappears entirely. Proof has the opposite properties: it is fixed when the work runs, it is complete rather than selected, and it can be recomputed by someone who does not trust the writer. A brief is for reading. A receipt is for checking. Confusing the two is how plausible summaries turn into wrong decisions.
How do I get receipts underneath my agent's briefs?
Route the work through a layer that receipts it as it runs. Add the Zambo endpoint (https://zambo.dev/mcp) as a connector in the AI you use, and the calls it makes there each return an AER-1 receipt at a public URL. No personal agent shipping today attaches portable receipts to its brief on its own, Duet, Muse, and Spark all summarize, so the connector is the current route. The brief keeps its job, orientation, and the record takes the other job, evidence.
What should I spot-check in a brief each morning?
One claim: the one with consequences. The payment it says went out, the message it says was sent, the booking with the tight connection. Open that claim's receipt and recompute the hash; a match means the saved result has not changed since the work ran. Then scan the brief for claims with no receipt anywhere beneath them and read those as plans, not facts. Two minutes, once a morning, and the brief becomes a table of contents for a record you can trust, instead of the record itself.