Preparing to Work Alongside AI Agents

In the previous post, I argued that meaningful human oversight requires more than placing a person somewhere in an automated process. The person must be able to see what the AI is doing, understand enough to evaluate it, intervene before harm occurs, and stop the system when necessary.

That becomes more difficult when AI moves from producing answers to taking actions. Most of our current conversations about student AI literacy still focus on chatbots. A student asks a question. The AI generates a response. The student decides what to do with it.

That interaction can create serious concerns about accuracy, bias, privacy, and overreliance. But the boundary between the human and the system remains relatively visible. The AI produces information, while the person acts on it. AI agents begin to blur that boundary.

An agent does not simply respond to a prompt. It may interpret a goal, break that goal into smaller tasks, select tools, retrieve information, make intermediate decisions, take actions, evaluate progress, and adjust its approach. Instead of asking an AI system to draft an email, a user may ask an agent to identify who needs to receive the message, gather the necessary information, prepare the email, attach the appropriate documents, send it, and schedule a follow-up. Instead of asking for troubleshooting advice, a technician may authorize an agent to examine logs, compare configurations, make changes, restart services, test the result, and document the resolution. Instead of asking AI to summarize a group of job applications, an organization may use an agent to retrieve applications, screen candidates, rank them against selection criteria, send interview invitations, and update the hiring system. Instead of producing a recommendation that a person must act on, the AI begins to carry out the work itself.

This is not simply a more powerful chatbot. It represents a shift in delegated authority. And that means preparing students to work alongside AI agents will require more than teaching them to write better prompts. They will need to learn how to define goals, establish boundaries, assign permissions, monitor progress, evaluate intermediate decisions, and determine when an agent should stop and return control to a human.

In other words, they will need to learn how to delegate.

Delegation is already an important professional skill. When we delegate work to another person, we do not simply state the desired outcome and walk away. Effective delegation usually involves deciding whether the person is appropriate for the task, explaining the purpose, identifying constraints, clarifying what authority they have, establishing checkpoints, and determining when they should ask for help.

We also consider the consequences of misunderstanding. If the task is routine, familiar, and easily corrected, we may provide broad direction and allow considerable independence. If the task is sensitive, unfamiliar, consequential, or difficult to reverse, we are likely to provide more specific instructions and closer supervision.

Many of these same principles apply to AI agents. But there is an important difference. A human colleague brings lived experience, social understanding, professional norms, and an ability to recognize that a situation has become unusual. They may notice that an instruction conflicts with the purpose of the work. They may question whether an action is appropriate even when it appears technically permitted.

An AI agent does not necessarily understand the larger meaning of the task. It may pursue the stated goal in ways the user did not anticipate. It may satisfy the wording of an instruction while violating its intent. It may make a locally reasonable decision that creates problems elsewhere in the workflow. It may continue acting when a human professional would pause.

This is sometimes described as the difference between specifying a goal and specifying the boundaries around that goal. Imagine telling an agent to reduce the number of unresolved customer-support tickets. The objective appears clear. But how should the agent pursue it? Should it send automated responses? Close old cases? Prioritize easy requests? Escalate difficult cases? Combine similar complaints? What counts as resolved? What information can it access? What promises can it make to customers? When must a person become involved?

Without carefully defined limits, the agent may achieve the measurable goal while undermining the reason the goal mattered. Closing tickets is not the same as helping customers. Scheduling more appointments is not the same as connecting people with appropriate services. Reducing the time required to evaluate applications is not the same as making better hiring decisions. Increasing student completion is not necessarily the same as supporting learning.

AI agents may be especially effective at optimizing what has been made explicit. Professional judgment is often required to protect what remains implicit. That is why one of the most important questions a graduate can ask is not only, “What do I want the agent to accomplish?” It is also, “What must not be sacrificed while it pursues that goal?”

This brings us back to the forms of thinking explored throughout the earlier series. Judgment is necessary to choose among possible actions. Sensemaking is necessary to understand how the task fits within the larger context. Ethical reasoning is necessary to consider who may be affected. Metacognition is necessary to recognize the limits of one’s own understanding. Discernment is necessary to decide which parts of the work should be delegated at all.

These human capabilities do not become less important when AI begins taking action. They become part of how authority is assigned.

Emerging research suggests that workers do not view all forms of AI delegation in the same way.

A large study involving 1,500 workers examined more than 800 tasks across 104 occupations. The researchers asked workers which tasks they wanted AI agents to automate or support and compared those preferences with assessments of current AI capabilities. Workers expressed different preferences for human involvement depending on the occupation and task. They did not simply want every technically automatable task to be automated. (Shao et al., 2025)

The fact that an AI agent can perform a task does not mean the people doing or affected by that work want it delegated. Workers may welcome AI assistance with repetitive documentation while resisting its involvement in interpersonal or identity-defining parts of their roles. They may want an agent to gather information but not make the final decision. They may want help preparing a communication but not have an AI system speak as though it were them. They may be comfortable delegating administrative coordination but not conflict resolution, evaluation, counseling, or relationship-building.

Technical capability is only one factor in responsible delegation. We must also consider consent, professional identity, workplace culture, trust, and the meaning of the task.

This becomes particularly important when an AI agent communicates on someone’s behalf. If an agent attends a meeting, responds to colleagues, communicates with clients, or represents a manager’s views, other people may not know whether they are interacting with the person or their automated delegate.

Research examining “manager clone” agents has begun exploring systems designed to reproduce a manager’s communication style and participate in meetings or exchanges on their behalf. These systems may reduce workload, but they also raise questions about authenticity, authority, surveillance, and whether employees can meaningfully distinguish the manager’s judgment from the agent’s output. (Hu et al., 2026)

A message sent by an agent may carry the authority of the person it represents even if that person never reviewed it. That creates a new professional responsibility. Graduates may need to decide not only what they are willing to let an agent do, but also what they are willing to let it do in their name. Would they allow an agent to communicate routine scheduling information? Respond to a customer complaint? Provide performance feedback? Negotiate a deadline? Decline a request? Represent their position in a meeting?

The answer should probably depend on the context, the relationship, the stakes, and whether the other person knows an agent is involved. Delegating the communication does not delegate its relational consequences. If an automated message damages trust, the human professional may still need to repair the relationship.

Agents also create new security and privacy concerns because action requires access. A chatbot can answer a question without connecting to other systems. An agent becomes more useful when it can access email, calendars, documents, databases, browsers, financial tools, customer records, or administrative platforms.

Each connection expands what the agent can accomplish. It also expands what can go wrong. An agent may retrieve information the user did not intend it to access. It may combine information across systems in inappropriate ways. It may follow malicious instructions hidden inside a document, email, or website. It may send information to the wrong person. It may make a technically permitted change that violates organizational policy. It may be granted broad access for convenience even though the task requires only a small portion of that authority.

Research on agent security has identified prompt injection as a particularly serious problem. An agent processing external content may encounter embedded instructions designed to redirect its behavior. Because the agent can use tools and take actions, a manipulated response may have consequences beyond producing incorrect text. (Abdelnabi and Bagdasarian, 2026)

This is another reason that working with agents cannot be reduced to prompt-writing skill. The person delegating the task needs to think about permissions. What information can the agent read? What systems can it access? What changes can it make? What actions require confirmation? What records will it create? What happens if it encounters conflicting instructions? How will the user know what it did? Can its actions be reversed?

The safest approach is usually not to give an agent every permission it might need eventually. It is to give it the minimum authority required for the current task.

This parallels the data-minimization principle explored earlier in the series. Just as an AI system should receive only the information necessary for a task, an AI agent should receive only the access and authority necessary to complete it. This is sometimes called the principle of least privilege. It has long been important in cybersecurity. Agentic AI makes it a form of everyday professional judgment.

Students should understand that convenience and control often pull in opposite directions. An agent with broad access can complete more work without interruption. An agent with limited access must pause and ask for approval. The first feels more efficient. The second may be safer, more transparent, and easier to supervise. The appropriate balance depends on the task.

An agent organizing a personal reading list may need little supervision. An agent sending external communications, modifying records, moving money, changing system configurations, or making decisions affecting other people should operate within much tighter boundaries.

This is where checkpoints become important. A well-designed agentic workflow does not necessarily require a person to approve every minor step. That would remove much of the value of using an agent. Instead, the workflow should identify consequential moments where human judgment is most valuable. Before information is disclosed. Before money is spent. Before a message is sent externally. Before a record is deleted or substantially changed. Before a decision affecting someone’s opportunities is finalized. Before the agent moves beyond the original scope of the task. The human should not supervise every keystroke. They should retain control over meaningful commitments.

Current agent systems do not always make those controls clear.

A 2026 analysis of 30 commercially available AI agents found substantial gaps in publicly available information about autonomy, safety evaluation, accountability, and how agents interact with the broader systems around them. (Kapoor et al., 2026) This is an emerging field. Capabilities are advancing faster than shared standards, organizational policies, and research evidence.

Our students may enter workplaces where agentic systems are introduced before anyone has fully determined how they should be governed. They may receive access to tools without adequate training. They may be encouraged to automate workflows without understanding the information moving through them. They may inherit agent configurations created by someone else. They may be asked to supervise multiple agents completing work at a speed no individual could independently verify. They may not even be told when an agent is acting within a system they use.

Preparing students for this environment requires more than demonstrating what agents can do. They need opportunities to practice deciding what agents should be allowed to do. A student could be given a professional workflow and asked to divide it between human and agent responsibilities. Which steps can the agent complete independently? Which require approval? What information does it need? What information should remain inaccessible? What conditions should cause it to stop? How will its actions be documented? Who will be notified if something goes wrong? How can the outcome be reversed?

Students could also examine how the same agent should operate differently depending on the context. An agent scheduling an informal team meeting might act independently. An agent scheduling a meeting involving confidential employee concerns might require direct human involvement. An agent drafting a routine service update might be permitted to prepare and send it. An agent responding to an angry customer or communicating about a serious failure might prepare options but leave the final response to a person.

The technology may be the same. The appropriate authority is not. This is the central lesson students will need: Using an AI agent is an act of delegation, and delegation is never neutral. It reflects what we value, what we trust, what we want to remain involved in, and what risks we are willing to accept.

AI agents may eventually become ordinary professional collaborators. But we should be careful with the word “collaborator.” An agent may perform tasks associated with a colleague without sharing the obligations, relationships, or consequences that human colleagues carry. It does not experience the workplace. It does not build a career. It does not feel responsibility. It does not become accountable simply because we give it a role.

The human and the organization remain responsible for defining the conditions under which it acts. Our graduates will need to be more than users of AI agents. They will need to become thoughtful delegators.

They will need to understand that a clear goal is not enough. Authority must have boundaries. Actions must remain visible. High-stakes commitments must remain under meaningful human control. And some parts of professional work may be too relational, consequential, or identity-defining to delegate at all.

That final idea leads to the last post in this series. As AI becomes capable of producing more of the visible work associated with a profession, students will need to decide what kind of professional they want to become and what they want their work to represent.

That is where the series will conclude.

Continuing the Conversation

Series 1: AI Is Exposing Existing Problems ✓ Completed
Series 2: What We Do About It ✓ Completed
Series 3: Cultivating Human Thinking in an AI World ✓ Completed
Series 4: Learning Alongside AI ✓ Completed
Series 5: Preparing Students for an AI World
Current Post (7 of 8): Preparing to Work Alongside AI Agents
Next Up: Building a Professional Identity in an AI World