You can supervise only as many active AI agent roles as your week has review time for. Starting a role is a one-time decision. Checking its outputs, repairing its handoffs and handling its exceptions are obligations that come back every week. Launching another role does not add an hour in which to judge its work.
In The Agent Dream Team, I make that constraint visible with a Roster Capacity Budget. My starting estimate is 3.75 attention units per role per week, where one unit is a focused half-hour. In the book’s illustrative 15-unit week, four roles use all 7.5 hours of supervision time, and a fifth fits only after review savings have been earned somewhere else. I counted my own week and rounded. The result is a planning tool, not a measured limit for AI agents, and no substitute for counting your own hours.
What are you counting: agents, chats or roles?
Count active roles with their own review obligation in your specialist roster. A Research Specialist and a Draft Editor are two roles even when they run on the same model: they have different mandates, outputs and scorecards. Five chat windows doing the same narrow job are not automatically five roles. The useful question is how many kinds of work you must route, inspect and decide on.
Include a coordinating role. In the book’s design, a coordinator that routes work between specialists is a roster role like every other: it carries a charter, consumes its own 3.75 units and sits on the same Trust Ladder. Delegating coordination does not delete your review duty; it relocates it. Leave out an experiment you have stopped and that can no longer receive work. Keep a role that is merely idle this week but still on the trigger map: the obligation returns the moment its trigger fires.
This is the distinction between launching and supervising. Writing a charter and connecting a tool get a role started. Keeping it useful means checking every week whether its trigger fired correctly, whether its output met the standard, whether its handoff carried the right context and whether the role should change or retire. ROSTER organises that work into six decisions per role, from Roles to Rotation, and the ROSTER explainer walks through them. The question here is narrower: can your calendar carry them?
How much weekly attention does one role cost?
Chapter 3 splits a typical role’s weekly supervision into four lines. They are my rounded planning estimates, not measured averages. Replace them with your own time record as soon as you have one.
| Management work per active role | Attention units per week | What you check |
|---|---|---|
| Charter maintenance | 1.00 | Mandate, boundaries and current objective. |
| Trigger monitoring | 0.50 | Work that fired, work that should have fired and work that should not have. |
| Output review | 2.00 | Delivered work against its rubric; handoffs and exceptions can add time. |
| Rotation decisions, averaged across weeks | 0.25 | Whether to improve, merge or retire the role. |
| Starting total | 3.75 | An estimate to test, not an allowance to spend blindly. |
At two units per supervision hour, required units = active roles × 3.75 and available units = protected weekly supervision hours × 2. Where your own record shows a different cost for a role, use that figure instead. The formula is a first pass, not a reason to force unlike roles into one box.
On the book’s example of 15 available units, three roles cost 11.25 and leave 3.75 for the week that goes wrong. Four cost 15 and leave nothing. Five cost 18.75, so at least 3.75 units must come out of review costs before the fifth fits. That arithmetic is what the book’s three-to-five working band describes. It does not license a fifth role whose checks are quietly skipped.
Output review is the largest line and the only one that scales with how much a role produces. If your outputs are long, technical or legally consequential, your review line is not 2.00 and your ceiling is lower than mine. Trigger monitoring is cheap until a trigger misfires. Rotation is amortised: in the book’s phrase, it costs you nothing for eleven weeks and then costs an afternoon. Record those overruns instead of reading the table as a cap on how long careful review may take.
The same chapter names the failure to watch for. When you exceed the budget, review quality degrades before output quality does. The outputs keep arriving and keep looking fine. What quietly stops is you reading them properly.
Is there a universal number of agents one person can manage?
No. Chapter 3 cites Gerald Bell’s 1967 study of supervisors and their employees in a community hospital as a human-team analogy. The paper’s abstract reports smaller spans of control when subordinates’ or supervisors’ jobs were more complex, and no relation between closeness of supervision and span. It does not measure AI agents and does not validate 3.75 units. The analogy points you at the complexity of the work and its review cost, not at a headcount borrowed from a hospital.
In the book I also call eight the practical ceiling: past eight, in my own roster and the ones I have watched, the founder stops being a coach and becomes the bottleneck again. It is a default for founders who review everything, not a limit measured on anyone. Do not aim for eight because the number exists. Three demanding roles can already exceed your week, while a larger roster of narrow roles fits only after review costs have genuinely fallen. Your time record decides what fits now.
Which hours belong in the capacity count?
Count the hours you can protect for supervision after reserving time for the work you still do yourself. Chapter 10 asks for the units you honestly have, not the hours you wish you had. An empty calendar block that client delivery will swallow is not capacity. Plan on a normal week, then check whether a busy week would break the plan.
For each active role, record the minutes spent on its charter and objective, trigger checks, output review, handoff repair and decisions about change. Include the minutes spent reading and approving work before a consequential action. If last week was unusually quiet, use a week with the role’s usual output, or mark the estimate provisional and time its next run.
The aim is to expose hidden review debt. If a scorecard is blank because nobody read the output, a low recorded time is not a saving. If a coordinator bundles five outputs into one handoff, you may save context switches, but you still have to check the packet and the work behind it. Count a saving only once your time record shows it.
Worked example: four roles and six hours
This operator and week are hypothetical. A solo founder can protect six hours a week for supervising agents: 12 attention units. The roster holds a Research Specialist, a Draft Editor, an Ops Coordinator and a GTM Distributor, all new enough to need the book’s full starting cost.
| Budget check | Units | Hours |
|---|---|---|
| Four roles × 3.75 units | 15 | 7.5 |
| Protected supervision time | 12 | 6 |
| Shortfall | 3 | 1.5 |
The founder pauses the Distributor, which handles one small batch a week, and postpones that batch while timing the other three roles. Three roles at the starting cost need 11.25 units. That leaves 0.75 unit, or 22.5 minutes, of the six hours. The arithmetic fits, barely.
Then the Research Specialist’s source checks take one unit more than estimated. The three-role total becomes 12.25 units, above the founder’s 12. Pausing the fourth role was necessary but not sufficient. The founder must protect more time, assign less work or change another role, then measure again. The budget exposes the gap; it cannot choose which work matters most. If some of that work could run as a script or a single model call, when not to use an AI agent helps you take it off the roster.
Can the 30-minute weekly ritual cover all review?
No. The book’s weekly operating ritual takes 30 minutes across the whole roster, with one ROSTER Field Card per role. On Monday, before the first agent session, you spend ten minutes checking each role’s objective for the week. Midweek, five minutes go to the trigger log: false positives, false negatives and delays. On Friday, fifteen minutes go to reading the scores recorded during the week, grading each weekly objective from 0.0 to 1.0 beside its score without averaging the two, and raising rotation flags. The rotation decision itself takes a separate fifteen-minute session. The ritual belongs inside the supervision budget. It is not the whole budget.
Chapter 8 adds two checks during the week: a 30-second direction check within the first 20 percent of a long task, and one to three minutes of rubric scoring per output, at delivery or in a review window within the same session cycle. That scoring is paid inside the 2.00-unit output-review line. Friday’s pass reads the recorded scores and looks for patterns; it does not reread the work. So a role can look cheap simply because nobody scored its outputs, and a weekly ritual cannot find what nobody recorded. Chapter 11 says what to do then: spot-check the most consequential output, log the gap as review debt and adjust the role’s cadence.
Some actions keep a founder gate whatever a role scores. In the book’s design, irreversible, public, financial and legally binding actions are capped at L3 on the Agent Trust Ladder, and the founder confirms them before execution, every time. A clean scorecard history can justify lighter review of routine work inside the charter. It never removes that gate. If an approval request can sit unanswered, the human approval timeout test shows why silence must never become approval.
What if the roster is over budget?
Chapter 10 gives three moves before anything new joins: drop the weakest addition, merge two low-frequency roles into one digest, or buy down review cost on a role with a clean scorecard history. Chapter 3 adds batching, so that five outputs cost one context switch instead of five. In the composite case that closes chapter 10, two quarters of clean scorecards move two roles to L4 (Ship-and-report: the role ships inside its charter and reports weekly), and that level takes a role’s output review from 2.00 units to 0.25 and its weekly cost from 3.75 to 2.00. Check the result of every move. A merger can make the combined charter harder to supervise, and a batch can hide a weak individual output if you stop looking at each one.
Do not raise a trust level to make the arithmetic fit. In chapter 8, trust is earned per role, never per model. Promotion takes four consecutive review cycles at or above the next level’s threshold, with no failure on a gate-relevant criterion. A charter change resets the role to L2 or below, and after a model change the new configuration has to requalify before it inherits the old level. If a role has stopped earning its place, retiring a production agent is a management decision, not a failed experiment. Close its triggers and revoke its access as part of the retirement, or the review obligation can return without appearing in your count.
Try this today (20 minutes)
- List roles (5 minutes). Write down every active role, including any coordinator, with its current output, trigger and reviewer.
- Name real capacity (5 minutes). Write the weekly supervision hours you can protect after your own delivery work. Multiply by two for attention units.
- Run the check (5 minutes). Multiply active roles by 3.75, and replace the estimate with measured minutes where you have them. Include output scoring, handoff repair and approval gates, not just the weekly ritual.
- Choose the next move (5 minutes). If required units exceed available units, name the role to pause, the low-frequency pair to test as one charter, or the review saving your scorecards already support. Then block next week’s review minutes in your calendar, so you can check the estimate against work that actually happened.
Write the result on one line: sum of role costs = required units; protected hours × 2 = available units; gap = available minus required. If the gap is negative, decide what stops before you start another role. If it is positive only because reviews were skipped, repair the review record first.
Cite this:How many AI agents can one person supervise?Len P. van der Hof. https://lenvanderhof.com/en/blog/how-many-agents-can-you-supervise/ ·