When employees supervise several AI agents, managers should limit concurrent work to what the team can check, interrupt and finish responsibly. Agree which requests need immediate attention, which can wait safely and who provides cover. Make room for supervision in the employee's existing job before adding more agents.
There is no defensible company-wide ratio in the evidence considered here. Three agents preparing drafts can create a different workload from one agent making time-sensitive changes. The useful question is what happens when several of them need the same person's judgment while that person is already busy.
Start with the employee's whole job
An agent can finish its part while creating another task for its supervisor. Someone must understand the result, resolve an exception or decide whether an action is permitted. Those demands can arrive during a customer call, a planning meeting or another agent's review.
In its June 2026 AI at Work survey, BCG reported that regular AI users could experience increased cognitive load alongside higher job satisfaction. The survey covered workers across 14 markets. It measured reported experience with AI broadly, not the safe number of agents for an employee to supervise.
That distinction matters for managers. Enjoying the tools does not establish spare attention. Nor does a growing list of completed agent tasks establish that the employee's overall job has improved.
Paul Boag described the operational friction in a May 2026 LinkedIn post: parallel agents can finish new tasks before the previous results have been reviewed. He advocated pacing and batching reviews. It is an adviser's observation, not a measured staffing rule, but the unfinished review is a recognizable piece of work to discuss with an employee.
For one representative busy period, identify:
- Attention required. What must the person inspect, decide or recover, and what evidence do they need?
- Waiting conditions. What happens while the person is unavailable, including any deadline or expiry?
- Competing duties. Which customer, colleague or operational responsibilities need the same person's attention?
- Coverage. Who else can take a decision, and what work must move to make that possible?
The guide to measuring checking burden explains how to record review effort, waits and unresolved results separately. Here, use that evidence to decide how much work should run at once.
Examine a collision before increasing concurrency
Suppose Sam Taylor coordinates a freight desk at a regional wholesaler. One agent prepares carrier exception proposals, another drafts customer delivery updates and a third prepares internal summaries. Sam also answers carrier calls and checks the day's dispatch plan.
At 14:50, the carrier agent flags a delivery postcode that conflicts with the order record. The dispatch cutoff is 15:00. Customer-update drafts become available at the same time, and the summary agent asks about a missing internal category. Sam is already on a carrier call.
Counting three agents tells the manager very little. The conflicting postcode needs the authorized dispatch person's decision before any affected release. The customer drafts can remain unsent. The internal summary can wait. A notification sound or the order in which requests arrived does not establish that priority.
Three requests compete for Sam’s attention
14:50 at the freight desk. Dispatch cutoff 15:00. Sam is on a carrier call.
Conflicting postcode
Keep the affected dispatch unreleased. An authorized deputy needs the order, proposal and current status.
Drafts ready to check
Keep them unsent while the dispatch decision takes priority. Review against the confirmed outcome later.
Missing category
Defer clarification and limit new starts. Confirm no essential downstream task depends on this summary.
Sam's manager should establish who handles the dispatch discrepancy and who covers the call. If an authorized deputy takes the discrepancy, they need the original order, carrier proposal and current dispatch status. Merely forwarding the agent's summary transfers a question without necessarily transferring enough context to answer it.
The waiting behavior also needs demonstration. Confirm that holding the customer drafts prevents sending and that the disputed dispatch cannot proceed without its required decision. If the cutoff passes unresolved, use the company's existing exception process. A deadline should not silently grant permission or create a delivery promise.
These are different controls from an agent's action-level approval boundary. That boundary says what may happen. The manager must also ensure someone can make the required decision when it is needed.
Protect the main task when routing requests
A shared view of pending work can help people see what needs attention. A coordinator agent may also collect status or assemble evidence. Neither arrangement establishes that the most important work will receive enough attention.
A useful caution comes from a 2012 experiment on scheduling operator attention. Thirty participants supervised eight simulated search-and-rescue robots under different queue arrangements. Showing all requests helped participants resolve more robot failures than a first-in-first-out queue, but they also missed more visible search targets. The study found no significant difference in the total number of targets identified.
This was a short laboratory robotics experiment, not a workplace AI trial or an eight-agent staffing recommendation. It does show why “more requests cleared” and “the main task performed well” deserve separate checks. For Sam, clearing summaries faster would not compensate for mishandling a dispatch discrepancy.
A coordinating agent creates its own work too. In a Reddit discussion about parallel coding sessions, a Reddit user described trying a manager agent and finding that supervision had shifted to that agent. Other replies described benefits from task boards and explicit handoffs. These are unverified personal accounts, not evidence that coordination always helps or always fails.
Before adopting another layer, check:
- Net effort. Does it reduce the person's work after its routing and summaries have been checked?
- Evidence access. Can the person inspect the originals behind a consequential request?
- Priority challenges. Can an authorized person change an incorrect recommendation?
An easier-to-read queue can still contain the wrong recommendation.
Change workload and coverage together
Do not ask an employee to supervise agents continuously while retaining a full schedule of duties that make them unavailable. Decide what the manager will remove, reschedule or cover.
The UK's Health and Safety Executive workload guidance recommends examining normal and peak conditions, involving workers in task analysis and considering how work is distributed across a team or shift. It is general human-factors guidance, rather than an agent-specific capacity formula.
For Sam's desk, distinguish coverage from priority:
| Staffing condition | Manager's decision | What to establish |
|---|---|---|
Sam has a protected review block | Assign cover for incoming calls | Another person has actually accepted the calls. |
Sam is on a call and a deputy is available | Route the dispatch decision to the deputy | The deputy has authority, original evidence and current status. |
Sam and the deputy are both occupied | Limit new time-sensitive work | Pending decisions stay held and the manual exception route has an owner. |
A decision remains open at shift end | Arrange an accepted handover or safe closure | The next person knows the deadline, waiting state and unresolved question. |
These are operating choices to test, not assumptions about what the software already supports. If the current tool cannot hold an action reliably, a review block is not an adequate control for it.
Ask the employee which switches are hardest to recover from and which other duties are slipping. Treat those answers as evidence about the arrangement. A person preferring fewer simultaneous tasks does not, by itself, show a training deficit. Better evidence packaging, less parallel work or a different schedule may be the useful change.
Establish a local limit through a bounded trial
Choose a defined task mix and a period that includes ordinary work and a plausible collision. Preserve the required checks and authority rules throughout the trial.
- Observe the current arrangement. Record demands on the employee, unresolved work, quality and the other duties they must complete.
- Rehearse competing requests. In an approved test or dry run, make the dispatch discrepancy and routine drafts arrive together. Demonstrate the hold and deputy handoff.
- Change one part of the load. Stagger starts, reduce parallel work or change coverage. Keep enough context to understand what changed.
- Review the whole result. Check whether required decisions happened in time, whether the ordinary job suffered and how demanding the employee found the arrangement.
Keep a locally tested limit attached to its conditions: task mix, work volume, available people, permitted actions and response needs. Reconsider it when those conditions change. A quiet week with experienced cover does not establish capacity for a busy week without that cover.
The tools-and-workflows guide also helps assess whether a simpler arrangement could meet the need. Reducing autonomy or running work sequentially can be a better operating decision than adding another supervisor layer.
Questions about supervising several agents
How many AI agents can one employee safely supervise?
The count depends on the work, its consequences and the employee's available attention and authority. A set of draft-only tasks can require different support from time-sensitive actions. Start with a bounded task mix, observe ordinary work and rehearse simultaneous requests. Keep the resulting limit conditional on coverage and workload, rather than applying another team's agent count.
Can a manager agent remove the need for supervision?
A coordinating agent may reduce status gathering or routing, but its priorities and summaries can also need checking. Compare the work it removes with the work it creates. Keep required human decisions and original evidence available, and test an incorrect routing recommendation before relying on the new arrangement.
What should happen when several agents need attention at once?
Prioritize by consequence, waiting conditions and the real response deadline. Hold actions that need an unavailable person's decision, route to an authorized deputy where one is available and limit new starts if the team cannot cope. Use the freight-desk collision to examine those decisions without assuming that an alert's arrival order defines importance.
Should every employee be expected to run agents in parallel?
No. Parallel operation needs evidence that the person can perform the required checks and their other duties under the actual conditions. Discuss difficult context switches and test fewer simultaneous tasks or adjusted coverage. Keep necessary reviews intact; a preference for sequential work is not evidence of poor performance or inadequate AI training.
How can a manager tell whether the arrangement is working?
Look at accepted work, missed or late decisions, unresolved requests, the employee's other duties and reported workload. Completed agent tasks alone are insufficient. If the queue clears while dispatch errors increase or the person cannot finish essential calls, change the arrangement. The bounded trial gives a practical starting point for that review.
