Supervising AI Agents: What Human Supervisory Control Can Teach Us About Agentic Work

The idea that humans will increasingly supervise AI agents can sound like an entirely new problem created by generative AI when, in fact, it isn't.

Human-factors researchers have spent decades studying what happens when technology becomes capable enough to perform work that a person previously performed directly. The terminology came from environments such as aviation, industrial process control, transportation, robotics, and military systems, but the underlying change was remarkably similar to the one now appearing in offices: the human stops executing every step and begins planning, directing, monitoring, and intervening in automated work.

That field is known as human supervisory control.

Its history matters because many of the problems now being discovered around AI agents were visible in earlier generations of automation: loss of situation awareness, reduced skill, inappropriate reliance, unclear function allocation, difficulty taking back control, and the uncomfortable reality that automation can make the remaining human work harder rather than easier.

AI agents bring those questions into a very different operating environment. The supervisor is no longer necessarily a pilot, plant operator, or robotics specialist. At any number of companies it could be an accountant, developer, analyst, procurement manager, marketer, security professional, or executive.

That makes supervisory control an important intellectual foundation for understanding human-agent work. Let's unpack it here.

What is human supervisory control?

Human supervisory control is a model of human-machine interaction in which automated systems perform parts of a task while a human sets goals, provides instructions, monitors performance, intervenes when necessary, and learns from the results.

Thomas B. Sheridan was one of the foundational researchers in the field. Supervisory-control research developed around the recognition that increasingly capable computers and automated systems were changing the role of the human operator. Rather than manipulating every control directly, the person increasingly provided higher-level objectives and constraints while automation performed lower-level activity.

This was already an established research problem by the 1970s. A 1976 NATO symposium edited by Sheridan and Gunnar Johannsen described supervisory control in complex human-machine systems where computers were beginning to share significant functions with human operators. Sheridan subsequently developed the model extensively across robotics, transportation, aviation, and other automated systems.

The basic supervisory-control model describes five human functions:

Plan → Teach → Monitor → Intervene → Learn

In Sheridan's formulation, the person develops a plan, communicates or programs that intent into the automated system, monitors execution, intervenes when performance diverges from what is required, and learns from the outcome to improve future interactions.

The technology was very different. The human problem is surprisingly familiar.

Knowledge workers are beginning to become supervisors of automation

Consider what happens when an employee delegates a meaningful task to an AI agent.

They decide what they want to accomplish, give the agent instructions and context, establish whatever constraints or access are necessary, observe what it produces or does, determine whether the work remains acceptable, and intervene when it does not. Ideally, what they learn from the outcome influences the next delegation.

The interface is different from an industrial control room, but the behavioral structure is familiar. A knowledge worker using an increasingly autonomous agent is beginning to perform a form of supervisory control.

That change is easy to underestimate because the transition may happen gradually. Someone begins by asking an AI system to draft an email or summarize a document. Later they ask it to research an issue, synthesize evidence, recommend an action, update a system, or execute part of a workflow. Eventually, the person may specify the desired outcome while the agent determines much of the process required to reach it.

The technology has moved from assisting the employee with execution toward performing work under human direction, and the employee's role changes with it.

This is part of the wider operating transition we explore in Human Resilience Engineering for Agentic AI. As agents perform more of the execution, human performance increasingly depends on whether people can delegate appropriately, maintain enough visibility, recognize meaningful deviation, and intervene effectively when conditions require it.

Automation changes human work rather than simply removing it

One of the most durable findings in the human-automation literature is also one of the most important for today's AI programs.

Raja Parasuraman, Thomas Sheridan, and Christopher Wickens made the point explicitly in their influential 2000 model of human interaction with automation. They argued that automation can operate across four broad functions:

Crucially, each function can exist at different levels of automation. A system might automate information gathering heavily while leaving decisions to the human, or it might analyze information, choose an action, and increasingly execute that action itself.

The authors' central observation was that automation changes human activity and can create new coordination demands for the human operator.

That distinction is highly relevant to AI agents because the governance problem changes as automation moves through those functions. An organization's risk profile is very different when AI retrieves information than when it analyzes that information. It changes again when AI recommends a decision and again when the system is permitted to execute one.

The workforce capability required at each stage changes as well. Someone supervising automated research needs to understand whether relevant evidence was missed. Someone supervising automated analysis needs enough expertise to challenge interpretation. Someone overseeing AI-supported decisions needs to understand when judgment should remain human, while someone supervising execution needs visibility and intervention mechanisms that still work after the system begins acting.

Talking about an employee simply “using AI” collapses all of those working relationships into one category.

For risk management, that is increasingly inadequate.

The old automation problem has arrived in the office

Classical supervisory-control research tended to focus on systems where both the stakes and the automation were obvious: aircraft, nuclear facilities, military systems, transportation networks, and industrial processes.

Generative AI brings a similar pattern into ordinary knowledge work.

A person may now delegate increasingly complex work through an interface that looks like familiar business software, which means the risk can develop without anyone thinking of the employee as a supervisor of automation.

A procurement professional might use an agent to gather supplier information, then evaluate submissions, then rank candidates, and eventually initiate routine follow-up actions. A software developer may move from code suggestions to an agent that investigates issues, modifies files, and opens pull requests. A financial analyst may begin with summarization and gradually rely on the system to perform much of the analysis that informs a recommendation.

The job title has not changed, but the distribution of work and judgment has.

Our Guide to Human-AI Risk Intelligence looks at why that distinction matters. Usage data can tell an organization that employees are using AI; it tells us much less about whether their roles are shifting from execution toward delegation, supervision, verification, and exception handling.

Those changes belong in the risk picture.

The irony of automation is still with us

In 1983, human-factors researcher Lisanne Bainbridge published a paper with a title that has aged extremely well: Ironies of Automation.

Her argument addressed a recurring problem in highly automated systems. Designers automate the tasks that technology can perform reliably while leaving people responsible for the unusual, difficult, or abnormal situations automation cannot handle. Yet those are precisely the moments when the human may be least prepared to take over because routine involvement in the process has disappeared.

Automation can therefore leave the human with less practice and a harder job.

That problem deserves renewed attention as organizations deploy AI agents. The operating model sounds attractive: agents perform routine work while humans handle the exceptions. The more difficult human-factors question is what happens to the human's ability to handle those exceptions after the agent has successfully performed the underlying task hundreds or thousands of times.

Expertise is maintained partly through doing, and context is maintained partly through participation. Situation awareness also develops through contact with the work. If automation removes enough of that contact, the human may still possess formal responsibility for intervention while becoming progressively less equipped to exercise it.

This is one reason effective human oversight cannot be reduced to placing an approval button at the end of a workflow. The person needs enough capability and context to know when approval would be a mistake.

The out-of-the-loop problem was identified long before AI agents

Research in the 1990s gave this problem another useful name: out-of-the-loop performance.

Mica Endsley and Esin Kiris studied what happens when people supervising automation lose sufficient involvement to respond effectively when the automated system fails. Their 1995 research linked automation with losses in situation awareness and slower recovery when participants needed to resume manual control. The paper discussed several contributing mechanisms, including loss of skill, vigilance problems, and the movement from active toward passive information processing.

An AI workforce can develop an analogous problem without any dramatic failure occurring.

Someone who once performed an analysis may become the reviewer of AI-generated analysis. Over time they spend less time constructing the reasoning themselves and more time evaluating a completed result. If the AI is consistently competent, review can become lighter and the person's detailed familiarity with the underlying work may decline.

Eventually, the organization can arrive at an uncomfortable configuration: the human remains responsible for recognizing when the system is wrong precisely because the system's success has reduced the human's opportunities to maintain the expertise required to recognize it.

That is a supervisory-control problem, and it is also a Human Risk Management problem.

Our work on Agentic Skill Atrophy explores the workforce version of this problem in more detail. The challenge is not simply whether people retain knowledge in the abstract, but whether they retain enough practical capability to supervise, challenge, and recover the work they remain accountable for.

Modern agent supervision adds new governance problems

Established supervisory-control research gives us an important foundation, but enterprise agentic work introduces conditions that earlier automation environments did not face in quite the same way.

Knowledge workers may delegate tasks informally and repeatedly rather than through a centrally engineered control system. Agent capability can change quickly, tool access can expand, and the boundary between drafting, recommending, deciding, and acting may move without a dramatic redesign of the workflow.

That creates a supervision problem that starts before the agent acts.

The person needs sufficient clarity about the objective, the task, the boundaries of delegation, the authority being granted, and the conditions under which the work should return to human control. Once the work begins, they need enough visibility to understand whether execution remains consistent with what they intended, without being buried under so much detail that meaningful supervision becomes impossible.

They also need to recognize when the work has moved outside expected conditions. That might be an unusual result, changing scope, an inappropriate assumption, an unexpected action, or a level of consequence that was not part of the original delegation.

Those situations are not always obviously right or wrong. New information may legitimately change the task, and the agent may occasionally identify a better approach than the one originally specified. Supervisory work therefore requires judgment rather than simple anomaly detection.

When intervention becomes necessary, the human also needs a usable means of exercising it. Correcting instructions, narrowing permissions, withholding approval, pausing execution, rolling back an action, or escalating the issue are only meaningful controls if the person has enough information and enough time to use them.

This is where modern agentic supervision begins to overlap with several of the concepts we use elsewhere in Human Risk Management. Intervention Points determine where meaningful human control can enter the workflow, while Intervention Authority determines what the person is actually allowed to do when they get there.

The principle is straightforward: a theoretical ability to intervene is not equivalent to practical control.

Delegation has to be revisited as the work changes

One of the more difficult characteristics of agentic work is that the original delegation may stop describing the work accurately.

An agent may gain additional access, begin handling a wider range of tasks, move further into decision-making, or simply become trusted enough that the employee stops supervising it in the way they did at the beginning.

The change may be gradual enough that nobody experiences a clear moment when the operating model changed.

Our Guide to Delegation Drift examines this directly. Delegated work, judgment, or authority can move away from what the person and organization originally intended without producing an obvious point at which somebody formally revised the arrangement.

Supervisory control therefore needs periodic reconsideration, not only monitoring of individual agent actions. The organization needs some way to notice when the objective, scope, system access, consequence, or degree of autonomy has changed enough that the original human role no longer makes sense.

The agent may still be working perfectly well while the supervision model around it has quietly become obsolete.

Technical agent controls do not solve the whole supervisory problem

Modern agent systems can contain sophisticated planning, monitoring, feedback, authorization, and runtime-control mechanisms of their own.

Agent platforms, IAM systems, and runtime security technologies may observe actions, enforce policies, constrain permissions, or stop prohibited behavior automatically. Those controls matter, particularly as agents gain access to systems and tools capable of producing real-world consequences.

They still answer a different set of questions from human supervisory control.

Technical controls can tell an organization what an agent was permitted to do and what it actually did. They cannot independently establish whether the person understood what had been delegated, recognized meaningful deviation, exercised appropriate judgment, or retained enough capability to intervene when the situation changed.

Our Guide to AI Agent Governance and the Missing Human Risk Layer explores that distinction in more detail. The technical control plane and the human supervisory system need to work together because each can see conditions the other may miss.

More autonomy changes the supervisory job

The Parasuraman-Sheridan-Wickens model also gives organizations a useful way to think about increasing agent autonomy.

Automation does not have one universal level. It can operate differently across information gathering, analysis, decision selection, and execution, which produces very different human jobs even when the employee remains “in the loop” in every version.

An AI system that retrieves documents mainly changes information acquisition. A system that synthesizes them moves into analysis. A system that recommends which supplier to select begins participating in decision selection. A system that then approves the supplier, updates the procurement system, or sends instructions begins participating in action implementation.

The employee may be present throughout the workflow, but their supervisory burden is not the same.

As automation moves further into analysis, decisions, and action, organizations need to revisit what the human must understand, how much attention supervision requires, where intervention can occur, and whether the person's retained skill remains appropriate for the responsibility they still hold.

That is why broad labels such as human in the loop can become misleading. They describe presence more easily than they describe supervisory effectiveness.

A better question is what the human is expected to do while they are there.

Supervision Sufficiency is the next question

Once an organization recognizes that employees are becoming supervisors of automation, another question follows naturally: is the available supervision actually enough for the autonomy and consequence involved?

Cybermaniacs uses Supervision Sufficiency to examine whether the human supervisory capacity around AI-enabled work is adequate for the task.

A person may technically occupy a supervisory role while lacking enough time, expertise, observability, intervention authority, or retained familiarity with the underlying work to perform that role effectively. The number of agents or workflows being supervised can also matter, particularly when employees are responsible for multiple automated processes while continuing to perform their ordinary jobs.

Supervisory control helps explain the job. Supervision Sufficiency asks whether the organization has made that job realistically performable.

That distinction becomes increasingly important as agent deployment scales.

AI readiness needs to prepare people for supervision, not just tool use

Many AI enablement programs still concentrate primarily on adoption: how to prompt effectively, use approved tools, protect sensitive data, and identify obvious limitations.

Supervisory work requires additional capabilities.

People need to understand appropriate delegation, establish useful constraints, recognize abnormal or consequential behavior, maintain enough skepticism to challenge the system without reflexively distrusting it, and know when a changed task requires another look at the authority already granted.

They also need enough domain capability to evaluate work that may increasingly arrive in finished form.

This is one reason our Agentic Readiness & Change (ARC) work looks at the workforce, roles, workflows, and governance surrounding agents rather than treating readiness as a training event. Agent adoption changes who performs the work, who supervises it, which decisions remain human, and what capabilities the remaining human role requires.

The technology can be ready while the supervisory system around it is not.

Organizations should design for supervision before supervision becomes the job

The history of automation gives AI leaders an unusual advantage: we do not have to discover every human-automation problem from scratch.

Decades of human-factors research have already shown that automation changes attention, skill, situation awareness, workload, reliance, coordination, and the nature of the remaining human work. It has also shown that the human retained as the final safeguard can become less capable of performing that role if system design removes too much meaningful participation.

Generative and agentic AI change the scale of the problem, the kinds of work involved, and the speed at which delegation can expand. They do not repeal the underlying human factors.

Organizations should therefore examine agent deployments as changes to the work system rather than simply as deployments of new software.

Who defines the objective? What does the agent actually perform? Which judgments remain human? What evidence does the person see? What constitutes meaningful deviation? When can they intervene? What happens when the task or authority changes? What capability does the supervisor need to retain? Who owns the outcome?

Those questions should be answered while the work is being designed, not after thousands of employees have quietly become supervisors of automation.

Human-agent work needs an engineering discipline

Supervisory control gives us an important intellectual ancestor for the emerging human-agent workplace because it helps explain why agentic AI cannot be governed exclusively as either a technology problem or an employee-training problem.

The relevant unit is increasingly the human-agent working system: the person, the agent, the task, the authority distributed between them, the information available to each, the environment in which they operate, and the mechanisms available when conditions change.

Human Resilience Engineering builds on that systems perspective.

The objective is not to keep humans artificially involved in every task agents could perform. It is to identify where human judgment, control, and adaptive capacity remain important and then design those responsibilities so people can actually perform them.

That may mean less human involvement in some parts of a workflow and much more deliberate involvement in others.

The research tradition behind supervisory control gives us several decades of evidence about why that design choice matters.

Agentic AI is making it everybody's problem.

Frequently Asked Questions

What is human supervisory control?

Human supervisory control is a form of human-machine interaction in which automated systems perform parts of a task while a human plans objectives, instructs the automation, monitors its performance, intervenes when necessary, and learns from the result. Thomas Sheridan and other human-factors researchers developed the field through work on increasingly automated systems beginning decades before generative AI.

How does supervisory control relate to AI agents?

People working with AI agents increasingly perform a similar role. Instead of directly executing every part of a task, a person may define an objective, delegate work, monitor the agent, evaluate its behavior, and intervene when necessary. This makes supervisory-control research highly relevant to contemporary human-agent work.

Who developed the theory of human supervisory control?

Thomas B. Sheridan was one of the foundational researchers in human supervisory control. His work described five central supervisory functions: planning, teaching the automation, monitoring, intervening, and learning from experience.

What are the main functions of supervisory control?

Sheridan's model identifies five major supervisory functions: Plan, Teach, Monitor, Intervene, and Learn. The model describes a feedback process in which humans establish objectives and instructions, observe automated execution, intervene when needed, and use experience to improve future control.

What is the out-of-the-loop problem in automation?

The out-of-the-loop performance problem occurs when automation reduces a person's active involvement enough that they lose situation awareness, skill, or readiness to resume effective control when automation fails or conditions change. Research by Endsley and Kiris demonstrated this problem in automated task environments, and the concept has clear relevance for people expected to oversee increasingly capable AI systems.

Why isn't “human in the loop” enough for AI governance?

Human presence does not guarantee effective oversight. The person also needs sufficient capability, context, authority, attention, and opportunity to recognize problems and intervene. As agents become more autonomous, organizations need to define the actual supervisory function rather than simply require a human approval step.

What skills do employees need to supervise AI agents?

Relevant capabilities can include defining appropriate objectives and boundaries, understanding what has been delegated, interpreting agent outputs, detecting meaningful deviation, questioning recommendations, controlling or stopping execution, escalating uncertainty, and retaining enough domain knowledge to evaluate the work.

The exact capability depends on the task, agent authority, and potential consequence.

What is Supervision Sufficiency?

Cybermaniacs uses Supervision Sufficiency to describe whether the human supervisory capacity around AI-enabled or agentic work is adequate for the autonomy, complexity, and consequence involved. It looks beyond the presence of a supervisor to whether that person has enough time, capability, visibility, authority, and retained familiarity with the work to supervise effectively.

How does supervisory control relate to Delegation Drift?

Supervisory control assumes that the human role and the automation relationship remain understandable. Delegation Drift becomes relevant when the practical scope of the agent's work, judgment, or authority gradually moves beyond what the organization originally intended, requiring the supervisory arrangement to be reconsidered.

How does supervisory control relate to Human Resilience Engineering?

Human Resilience Engineering applies a systems approach to the human contribution in AI-enabled and agentic work. Supervisory control provides an important foundation because it treats people and automation as parts of a joint operating system rather than examining human behavior or automation independently.