A weekly agent review should end with a decision for every active role, supported by evidence from its work. Decide what continues, what changes and what needs a separate investigation. Put the reason beside the decision so you can check it next week.
Chapter 11 of The Agent Dream Team gives this review a bounded shape: ten minutes on Monday, five midweek and fifteen on Friday. It calls the cadence the weekly operating ritual. The timings are a starting design to test on your own roster, not a proven optimum or the full cost of supervising agents.
Why isn’t a daily agent check-in enough?
A standup aligns teammates who must coordinate their day. A one-person company has no teammates to sync. The solo version is a founder opening one catch-all chat each morning to ask an assistant or agent roster what ran yesterday, what comes next and what is blocked. Chapter 11 describes where that leads: the daily check-in becomes another mega-thread, and context from early in the week contaminates the work at the end of it. Meanwhile the management question stays open: should each role keep running with its current objectives, instructions and permissions?
A completed task may answer the wrong question. A quiet role may have missed a brief. A busy one may keep producing work you have to repair. An activity summary alone does not settle any of those cases.
The distinction is reporting activity versus deciding what the evidence warrants. Agent roles still need operational coordination: inputs must arrive, handoffs must work and exceptions must reach you. Inspect those records. The weekly review gives you a fixed place to decide what they mean for the roster.
Keeping a role unchanged is a valid result, if the record says why its current work and permissions still fit. A review needs a decision you can account for, not a change.
How should a solo founder review their AI agents each week?
Use three passes, each with one question and a written result. Chapter 11 sets the starting point at two or more active roles with charters, triggers and scorecards in place.
| Pass | Timebox | Question | Record before closing |
|---|---|---|---|
| Charter review (Monday, before the first agent session) | 10 minutes | Are the objectives right for this week? | Confirmed or updated objectives, with design problems flagged separately |
| Trigger check (Wednesday, midday or early afternoon; Tuesday in a short week) | 5 minutes | Did qualifying events activate the right roles? | False positives, false negatives, activation delays and midweek objective updates |
| Scorecard pass (Friday, at the end of the working day) | 15 minutes | What does the week’s output evidence justify? | Scores, objective grades, card updates, rotation flags and next week’s objectives |
Monday directs the work before it starts, Wednesday checks routing while you can still respond, and Friday reads the accumulated evidence.
Keep the budget honest. Delivery scoring happens as outputs arrive (one to three minutes per output in Chapter 8); repairs and rotation sessions need their own time. Chapter 11 reserves 45–60 minutes for the first run as setup cost. From week two, return to the thirty-minute constraint and treat repeated overruns as a design signal.
What should be ready before the timer starts?
Prepare one ROSTER Field Card per role. Copy these eight fields from the documents you already use:
- Front: role name, one-sentence mandate, trigger and trust level (L1 to L5, meaning its permitted autonomy).
- Back: output contract, permitted tools, escalation rule and current scorecard grade.
The card gives you the current facts at a glance. It does not replace the charter or grant permission. Chapter 11 puts a first deck for three roles at about fifteen minutes: you copy fields rather than compose them. If a source field is missing, you have found a gap in the charter; record it as setup work. Do not invent an escalation rule to finish the card.
Keep the output log, delivery scores and trigger log within reach, plus a separate checklist for the week. Cards hold current state; checklists preserve history. Update a card in place when the role changes and keep each completed checklist so you can reconstruct why. When a role is retired, merged or migrated, move its card to an archive with the date it left the roster and one line naming the cause, so you do not rebuild the same failed design later under a new name.
A role with no trigger event this week can sit out the review. First confirm that no qualifying work arrived. A missed activation is a failure to investigate, not an inactive role to omit.
What happens in each pass?
Monday: confirm the objective before starting work
Read each card’s front and the role’s weekly objective. Mark the objective unchanged, updated or flagged. Check it against current deadlines and priorities, including last Friday’s findings.
Write a two-sentence revision where needed. If the problem requires a role redesign, schedule that work separately. Settle each objective before that role’s first session opens, and keep the hard stop: if you are still reviewing at minute ten, flag the remaining item and stop.
Midweek: compare incoming work with actual runs
Read the trigger on each card against incoming events and the trigger log. Chapter 11 scans for three misfire patterns: false positives (a role activated on an event that did not require its mandate), false negatives (qualifying events that started nothing) and activation delays (the right role started later than its trigger allowed).
Successful runs cannot show work that never started, so compare incoming briefs and requests with the activations. Record the event, expected role, actual result and misfire type. The trigger-map article covers designing and testing a route; this pass identifies where attention is needed.
Update an objective if priorities have changed since Monday. Handle urgent failures when they appear. The calendar sets review appointments, not waiting periods for known problems.
Friday: read the scores and write the decision
Start with scores recorded at delivery and open an output only to resolve a gap. Review it against its rubric criteria; leave line-editing for a separate session.
Keep the weekly objective grade (0.0 to 1.0, the share of Monday’s objective criteria met) beside the output-quality score (1 to 3 on the role’s rubric). One asks whether the role hit its target; the other asks how well the work met its contract. Never average them together.
For each role, apply its existing permission rules, note any charter problem or rotation flag, and write next week’s objective. Update the Field Card that evening: the new grade goes on the back. If the week’s average dropped the role below its level’s floor, or completed the fourth consecutive cycle a promotion needs, the trust level on the front changes too. The agent-scorecard article covers the criteria, release gates and permission decision. A high weekly average never clears a failed release gate.
Missing delivery scores are review debt. Chapter 11 says to spot-check the most consequential output, record the gap and adjust the role’s cadence before the next week. Keep the gap visible: one checked output cannot stand in for the whole week. Do not promote a role on evidence you have not collected.
When does the review trigger a rotation decision?
Chapter 9 separates a bad cycle from a repeated pattern. One week below the role’s floor calls for diagnosis. The book’s default rotation flag is two consecutive cycles below the floor, or four of the last six. These are the author’s operating rules, not validated reliability thresholds.
Friday raises the flag and books a separate 15-minute rotation session, ideally later that day or before Monday’s charter review. That session decides whether to retire, merge or upgrade. Check whether the model or platform changed before attributing the problem to the role’s design.
The separation protects coverage: one role’s investigation must not consume the time reserved for every other role. Record which outputs and recurring defect the session will examine, and when.
If the decision is retirement, the production-agent retirement checklist covers closing its triggers, keys, callers and recurring costs. A rotation flag retires nothing. Nor does a scheduled session justify postponing a permission restriction that a failure requires now.
Worked example: one hypothetical week
Consider a hypothetical solo consultant with three roles: Research Specialist, Draft Editor and Ops Coordinator. All remain draft-only (L1): the founder reviews every output before handoff. The work, scores and decisions below are invented to demonstrate the method. They are not a case from the book or observed results.
On Monday, a client proposal takes priority. The research objective changes from a broad market overview to: “For Thursday’s proposal, compare two delivery options against the client’s three requirements. Give a source or mark an unknown for each comparison.” The Draft Editor’s objective names the proposal’s required sections. The Ops Coordinator prepares the internal list of open actions. All three objectives remain within their existing mandates.
Wednesday’s check finds a new brief with no matching activation. The founder logs the false negative, handles the handoff manually after the check and schedules the routing repair. A list of completed agent runs would not have exposed that brief.
On Friday, the founder reads the delivery records:
| Role | Hypothetical evidence | Decision and next check |
|---|---|---|
| Research Specialist | The comparison includes broad industry background instead of answering the requirements. Recorded weekly quality averages are 2.3 last week and 2.4 this week, below this example’s 2.5 floor. | Keep draft-only supervision and book a separate rotation session after the Friday pass to inspect the repeated scope failure. |
| Draft Editor | The drafts contain the required sections, their claims trace to the research notes, and all recorded release conditions pass. | Keep the current trust level and objective pattern. Check the next deliveries against the same criteria. |
| Ops Coordinator | The action list includes completed jobs alongside open ones. | Change next week’s objective to list open actions only. Check that condition at the next delivery. |
The separate rotation session finds no model or platform change. The research mandate still fits the work, but the standing prompt still asks for broad background. The founder chooses a prompt upgrade: rewrite it from the charter, keep direct review and set a two-week re-evaluation date. That date is a checkpoint, not a promise of promotion.
The rotation log records the decision and its evidence now. Its outcome field stays open until later outputs show what happened. Nothing in this week yet demonstrates improved performance.
What if the review will not fit?
A first-run overrun is not a roster-size verdict: missing cards and scattered records eat time before you judge any work. Fix that preparation first.
Chapter 11 names three causes when the review keeps running long: too many roles, vague criteria and line-editing during evaluation. Keep rewriting outside the review. If you cannot score a criterion quickly, the chapter’s fallback is a good-or-bad judgment for that week: leave the criterion out of the average, hold any promotion that week and sharpen the wording on Monday.
If the active roster still exceeds the timebox once those problems are fixed, it is too large for the review time you have allocated. Reduce it; do not make the review fit by skipping the roles with the weakest evidence.
Try this today
Use your latest completed working week for this fifteen-minute test of the Friday pass. It adapts the book’s checklist; it does not replace running all three passes this week.
- Book the recurring appointments: ten minutes on Monday, five midweek and fifteen on Friday. Starting midweek? Confirm current objectives before the next agent session and pick up the remaining checkpoints.
- Open every active role’s output log and scorecard. Put its Field Card beside them, preparing missing cards first and timing that separately. If no scores were recorded, mark the evidence gap before starting.
- Set a fifteen-minute timer and write one decision per role. Record the output or score behind it, what happens next and when you will check again. Keep rotation investigations outside this pass.
- Make missing evidence actionable. Name the outputs still to review, when you will review them and which restriction applies meanwhile. An empty score column cannot justify more autonomy.
- Check coverage when the timer ends. Every active role needs a recorded decision, supporting evidence or a named gap, and a next check. Record unfinished roles explicitly and book any rotation sessions separately.
The test is whether you can account for every active role within the timebox without hiding missing evidence or doing the repairs inside it. If it does not fit, record what consumed the time and fix that friction. If it still will not fit, the roster is too big for one operator: take a role out of the active roster and take back its pending work yourself. A smaller roster you can review is the one you can justify running.
Cite this:Weekly agent review vs standup: one decision per role, not a status update.Len P. van der Hof. https://lenvanderhof.com/en/blog/weekly-agent-review-vs-standup/ ·