Direct answer

The week, reviewed from the record instead of from memory.

Every Friday the same meeting happens: each person and each agent reports the polished version of the week, and nobody can check any of it. The fix is not a better status template. It is reviewing a record that was written while the work ran, step by step, with the receipts underneath. Here is the whole review, in the order to run it.

See the instrument first: the live demo runs one real call and hands you its verifiable receipt, no signup.

Updated October 17, 2026. The two delivery records quoted below are real and public, with their client decision states read back from the live service on publication day.

Why the summary meeting cannot work

A weekly review built on summaries has three faults, and none of them are anyone's dishonesty. Summaries are written from memory, so the retries, dead ends, and quiet failures fall out. Each AI and each person can only report their own slice, so the week as a whole exists nowhere. And a summary reads exactly as clean whether the work went well or sideways, so the meeting cannot tell a good week from a lucky one. The review needs a source that was not written for the review.

The review, step by step

Step 1. Open the record by day. When the team's agent work runs through Zambo, every tool call saves a verifiable receipt at the moment it executes, and the day view at zambo.dev/today lays the day out in order. Walk the five days in sequence. You are not reading reports about the work; you are reading the work's own trail.

Step 2. Match shipped work against the record. For each thing the team says shipped, find the job in the record and open a step or two down to the receipt. Most weeks this takes minutes, because most claims check out, and the one that does not is exactly the one the meeting exists to find.

Step 3. Hunt the loops. Look for jobs that retried the same step repeatedly or burned calls and produced nothing. That pattern is next week's waste already scheduled. One prompt fix found this way pays for every review after it.

Step 4. Read the spend next to the work. The spend record rides the same connection, so cost questions are answered beside the steps they bought, not from a separate bill nobody can map to jobs.

Step 5. Close on what is still waiting. Finish with the handovers nobody has decided on. A finished job whose acceptance state still reads pending is a payment argument warming up. The review is where it gets chased while chasing is still cheap.

What the two states look like in a real record

Accepted: a job the review can close

Scope: Retrieval, extraction, verification, and reporting pipeline producing an AI-adoption research packet for a law firm.

Status: accepted by the client on 2026-10-16, decision ok_to_release.
The record covers 4 steps, and 4 of 4 steps were checked.
The acceptance receipt is 14cb8b21-d163-4c46-a692-b4947e3719ca, recorded against the record itself.
Open the handover or read its acceptance.json.

Still open: a job the review should chase

Scope: AI visibility audit of zambo.dev: homepage copy inventory, on-page citation audit, and evidence-based recommendations.

Status: still pending, decision hold. The client has not accepted or flagged it yet.
The record covers 5 steps, all saved on 2026-10-16 between 14:33:42 and 14:34:01 UTC.
Its not_checked list names two steps, aa725980 and 75253fbc, with the reason recorded: tool did not complete successfully.
Open the handover or read its acceptance.json.

Read together, the pair is the review in miniature. One job is closed, with the client's decision recorded. One job is finished but undecided, and its own record names the two steps that were not checked. No summary would have volunteered either fact. The record states both, because it was written during the work, not for the meeting.

What the record does not settle

A verifiable receipt proves the saved result was not changed since it was saved. It does not prove the work was good, it does not judge priorities, and it does not cover work that ran outside the connection. A weekly review still needs human judgment about what mattered. What the record changes is the floor of the conversation: nobody has to argue about what happened, so the hour goes to what it means and what to fix. AER-1, the receipt format, is an IETF Internet-Draft (draft-zambo-aer1-23), not a finalized standard.

Frequently asked questions

How do I review what my team's AI agents did this week?

Review the record, not the summaries. If the week's agent work ran through one connection, open the day view for each day and work through the week in order: every job, every step, the time each step was saved, each step carrying its verifiable receipt. Check four things. What shipped, matched against the record instead of against status updates. What looped, retried, or burned calls and produced nothing. What it cost, from the spend record on the same connection. And what is still waiting, meaning handovers whose acceptance state still reads pending and hold. One hour with the record replaces a week of polished status reports, because the record was written while the work ran, not composed for the meeting.

Why not just collect status updates from the team?

Because a status update is written by the person or agent whose work is being reviewed, from memory, for an audience. It rounds off the retries, drops the dead ends, and reads the same whether the week went well or sideways. The record has the opposite shape: it was saved step by step at execution time, before anyone knew there would be a review, and each line can be opened down to the receipt underneath it. Use updates for plans and judgment calls. Use the record for what actually happened. When the two disagree, believe the timestamps.

What should I look for in the weekly review?

Four patterns pay for the hour. Loops: a job that retried the same step again and again is a prompt or a tool that needs fixing, and it will waste the same hour next week. Silent failures: steps that were not checked or did not complete, named in the record itself, including the not_checked list a handover carries. Spend drift: calls that grew day over day on the same connection, visible next to the work they bought. And pending acceptances: finished work still sitting at hold because nobody on the client side has accepted or flagged it, which is next week's payment argument warming up. Pick the worst pattern, fix that one thing, and let the next week's record grade the fix.

Can I review work from different AIs in one place?

Yes, if the work ran through the same endpoint. A team's week usually scatters across several AIs and several people, and each chat keeps only its own trail. When the calls run through Zambo's one MCP endpoint, every step lands on the same record with its receipt, whichever AI ran it, and the day view at zambo.dev/today shows the saved record of that work in order. Each handover also names, per step, the AI label recorded for that step, so who did what is a line in the record rather than a reconstruction. Work that ran outside the connection is not in the record, and the review should say so plainly.

How much does a weekly review setup cost?

The record itself is a byproduct of running the work through the connection; the review is you reading it. For $0.99, put one full review day of team agent work on the record and run the Friday review from it: get the Day Pass. A Day Pass covers 24 hours, no subscription. Ongoing use is the Zambo Pass at $29 a month or $290 a year, and $19 a month for $ZAMBO holders, which adds private handover links, your branding, and the approval record as PDF and JSON files.