Supervision load and oversight instrumentation

Target state

In short: Supervision is designed like a queue with measured capacity, every agent has a named accountable person, and a monitor watches the monitors.

Supervision is engineered as a queueing system, not asserted as a role. Every agentAgent. Software that uses an AI model to work toward a goal on its own: it reads, decides the next step, uses tools, and keeps going until the job is done or a limit stops it. has a named accountable sponsorSponsor and owner. The sponsor is the business person accountable for why an agent exists. The owner is the technical person who keeps it configured and secure. recorded on its identityAgent identity. A registered account for an agent with a named owner, limited permissions, and credentials that expire, so every action can be traced to who asked.. Its credentials are scoped, expire, and require a fresh approval cycle. Sponsorship transfers automatically on departure rather than lapsing. Supervision is integrated with the work rather than handed to a specialist function. Oversight is designed for prevention and legibility, with verification made cheap by construction rather than demanded by policy. Supervision load is measured and capped by queueing behaviour rather than by agent count. A monitoring layerLayer (the fourteen layers). One of fourteen parts of the enterprise estate that agents touch, from infrastructure (R01) to agent data engineering (R14). watches the monitors: an algorithmic pass flags elevated-risk sessions before a human looks at anything.

layer · architecture

Supervision as a queueing system

Supervisory capacity is neglect time over interaction time plus wait time; load is budgeted as a burst rate per ten minutes, never a daily average.

1. 01Agent

   Agent

   Agent fleet

2. 02Control

   Control

   Algorithmic monitor

   Flags elevated-risk sessions; alarms on alarms

3. 03System

   System

   Escalation queue

   Wait time is the binding term

4. 04Human

   Human

   Supervisor

   Capacity = NT / (IT + WT) + 1

5. 05Evidence

   Evidence

   Supervision telemetry

   Intervention rate, escalation mix by trigger, wait per item

No credible published human-to-agent supervision ratio exists; burst bands replace ratios.

Diagram description: Supervision architecture from agent fleet through algorithmic monitoring and the escalation queue to the supervisor, with wait time explicit in the capacity model and telemetry feeding back. The map contains Agent fleet; Algorithmic monitor: Flags elevated-risk sessions; alarms on alarms; Escalation queue: Wait time is the binding term; Supervisor: Capacity = NT / (IT + WT) + 1; Supervision telemetry: Intervention rate, escalation mix by trigger, wait per item. Its connections are fleet to monitor; monitor to queue; queue to supervisor; supervisor to telemetry; telemetry to monitor for tunes flagging. Important boundary: No credible published human-to-agent supervision ratio exists; burst bands replace ratios.

Reviewed 2026-08-20·Sources:r13-operating-model findings

Mechanisms

Capacity is a queueing calculation, and wait time is the term everyone drops

In short: Capacity depends on how long items wait; ignoring wait time overstates it badly.

Supervisory capacity follows the fan-outFan-out. How many agents or tasks one supervisor can watch at once, which shrinks as waiting times grow. relation: neglect time divided by the sum of interaction time and wait time, plus one. The relation was measured across 2, 4, 6, and 8 supervised units. Interaction time stayed nearly flat (18.19 seconds to 15.74 seconds). Wait time grew roughly eightfold (8.71 seconds to 67.58 seconds). There was a statistically significant performance drop between four and six units, and at eight units each unit spent about half its time waiting. Ignoring wait time overstates capacity by up to 67 percent, and by 36 percent even in exception-based designs. Deliberate slack beats maximum fan-out: two operators each supervising five units outperformed one operator supervising ten, even carrying a 20 percent communication overhead.

Load is budgeted as a burst rate, never a daily average

In short: Plan for the busiest ten minutes; a daily average hides the moments when people are overwhelmed.

The only standards-embedded workload numbers come from alarm management, and they are written as burst ratesBurst rate. Demand measured at its busiest moments rather than averaged over a day. Oversight capacity is planned against bursts., not daily averages. The steady-state bands per operator are as follows. Fewer than one item per ten minutes is very likely acceptable. One per five minutes is manageable. One per two minutes is likely over-demanding. More than one per minute is very likely unacceptable. In the first ten minutes after a major upset, more than one hundred items is described as excessive and very likely to lead the operator to abandon the system. The governing rule comes from the same standards body. A per-ten-minute rate cannot validly be converted into a per-hour or per-day figure. The per-day metric was removed from the standard precisely because averaging destroys its meaning. Any supervision target expressed as a daily average will be met on the dashboard and violated in every burst. Consequence-weighted domains regulate below one-to-one: nuclear control rooms are required to staff multiple licensed operators per unit, with a licensed operator present at the controls at all times.

The telemetry that makes oversight evidence rather than assertion

In short: Measure four things about supervision itself, and check them against reality, because people misjudge their own speed.

Instrument four things, and read the telemetryTelemetry. The stream of measurements a system emits about itself, collected centrally. together. First, intervention rate and trend, understanding that a falling rate may be fatigue rather than improving reliability. Second, escalation mix by trigger, separating human-initiated from algorithm-triggered, because in the largest field deployment their outcomes differed materially. Third, wait time per supervised item, not agent count, because wait is what binds capacity. Fourth, verification cost per review, because that is the design lever. Add periodic blind checks, because self-report is unreliable in a specific and measured direction. In one study participants forecast being 24 percent faster, estimated afterwards that they had been 20 percent faster, and were measured 19 percent slower.

Verification cost is the lever; exhortation is not

In short: People check agent output when checking is easy, not when told to, so design the easy check.

One study covered roughly 55,000 agent-generated code-review comments in 342 repositories. The strongest predictor that a human acted on a comment was that it carried an inline suggestion. Long, complex comments were less likely to be acted on. Automation bias tracks verification complexity and cannot be prevented by training or instructions. Cognitive forcing functions (design features that make a person pause before accepting an answer) moved acceptance of wrong output from 0.64 to 0.48. The designs that worked best were rated most difficult, least preferred, and least trusted by the people using them. Design accordingly: make the correct check the cheapest available action, and expect the effective design to be unpopular.

Calibrated oversight, not a ratchet

In short: As trust grows, people should step in more often while granting more standing freedom; fewer interventions is the warning sign.

This is calibrated oversightCalibrated oversight. As operators gain experience they grant broader standing permission and also step in more often, rather than supervising less.: trust does not reduce supervision; it redistributes it. In measured deployments, auto-approval rose from roughly 20 percent to above 40 percent with experience. Interrupt rates rose from roughly 5 percent to roughly 9 percent over the same period. Rising interventions alongside expanding standing permission is the system working. Falling interventions alongside expanding autonomyAutonomy level (A0 to A5). How much an agent may do without a person in the loop, from manual (A0) to governed lights-out operation (A5). Set per task, not per agent. is the pattern to investigate. Named accountability is not itself a control. In testing, a phished employee's promptPrompt. The instructions and information given to the model for a task. A prompt is advice to the model, not a control on it. asked an agent to read cloud credentials and post them externally. It succeeded in 24 of 25 attempts, and the sponsor's name on the agent prevented none of them.

Design decisions

• Agents as workforce vs agents as tools (CD-22, whether to treat agents like staff members or like tools): keep the accountability, drop the personnel metaphor. The strongest field deployment achieves per-agent named accountability with no org-chart presence. Integrated supervisors outperformed specialists.
• The oversight gate for autonomy levels A4 and A5 is this layer's output. A4 and A5 are the two highest levels of independent operation, defined on the autonomy contract. The output is capacity calculated with wait time, expressed as a burst rate, and instrumented in production.

Cross-cutting concerns

Evidence and limits

The fan-out measurements, the alarm-standard bands, the code-review study, and the cognitive-forcing results are peer-reviewed or standards-published. The auto-approval and interrupt trends come from platform telemetry [vendor]. The central refusal: no credible published human-to-agent supervision ratio exists from any source. A major vendor introduced the metric as a management concept and published no figure in either of two annual editions. Every number in circulation is an extrapolation. Several widely repeated figures were traced to untraceable or misquoted origins and are excluded from this guide entirely. Design to the burst-rate bands and your own measured wait times instead. Two items to re-verify. First, whether any credible measurement of agent-supervision capacity is published. Second, whether the 2026 finding holds that the vigilance decrement (the fall in attention over a long watch) is a bias shift rather than a sensitivity loss. If it holds, it would redirect interventions aimed at improving detection.

The research behind this page

• Operating model findings
• Research brief
• Vendor landscape
• Sources