Every Knowledge Worker Is Becoming an Agent Manager

The role is changing

Role drift is a well-known problem with LLMs. If you have ever tried to make one play a specific human role throughout a long conversation, you have probably seen it happen. Sooner or later, it forgets the character it was supposed to play and drifts back into being a helpful assistant.

Humans tend to be better at sticking to the roles we are given. Or at least, we used to be.

If you are a software developer adopting AI as quickly as it evolves, you may have noticed that your role has quietly started changing. You still write code, but you increasingly spend your time describing tasks, supplying context, reviewing results, correcting mistakes, and coordinating work across several agents.

It did not happen gradually enough for us to properly prepare for it, yet it happened quietly enough that we barely noticed. The LLM may have drifted back into being an assistant, but you have been role-drifting too.

You are no longer only a software developer. You are becoming the manager of a small, fast, occasionally brilliant, and frequently confused team of agents.

This change is not limited to developers. Salespeople use agents to research accounts, prepare outreach, and develop leads. HR teams use them to summarize applications and prepare interviews. Designers direct agents through rounds of prototyping and experimentation. Product managers use them to analyze competitors, identify gaps, and challenge early ideas.

Across all of these roles, people are moving from performing every step themselves to directing, reviewing, and integrating work produced by AI. We have all stepped up a level in some sense. Each of us is becoming a team—and the job is no longer simply to use AI, but to manage how that team works.

The work of an agent manager

The new job is not to send an agent a sentence and hope for a miracle. It is to give direction, create the conditions for good work, and take responsibility for what comes back.

A good agent manager has to do several things well: choose the right work to delegate, explain the desired outcome, provide enough context, set boundaries, review the result, and decide what should happen next.

That sounds a lot like management because it is management. The difference is that an agent can move very quickly, produce a very plausible answer, and still have no idea whether the work matters.

An agent has no business priority of its own. It spends its capacity on whatever work it is directed, enabled, or allowed to do. If the objective is vague, it may confidently solve the wrong problem. If important context is missing, it may reconstruct information the organization already has. If its boundaries are unclear, it may go further than anyone intended.

The manager therefore remains responsible for the assignment. Not for personally performing every step, but for making sure the work has a purpose, the agent has what it needs, and the result is judged against something real.

The management loop

The first step is to start with meaningful work. What are we trying to achieve? What would a useful result look like? What should the agent not do?

The second step is to give the agent the context and authority it needs. Good managers do not just issue orders; they make sure their team can act. That may mean supplying background, examples, relevant documents, tools, decision boundaries, and a clear point at which a human must review the work.

The third step is to review the result. The agent may have completed the assignment, but that does not mean the assignment was completed well. Review is where we test assumptions, compare the output with the intended outcome, and decide whether to accept, correct, or reject it.

The fourth step is to learn from the pattern. Did the agent repeatedly need the same clarification? Did different agents fail in the same way? Did the work become easier once the context was improved? Did the manager keep delegating tasks that should have stayed with a human?

The final step is to improve the system. Refine the instructions, change the context, choose a better tool, adjust the workflow, add a checkpoint, or change the kind of work being delegated. Then run the loop again.

How to get better at the new job

The fastest way to become a better agent manager is to stop treating every failure as an isolated failure. A poor result is not only something to fix; it is information about the way the work was assigned.

When an agent goes down the wrong path, ask what it was missing. Was the goal unclear? Was the relevant context unavailable? Did the agent have too much freedom, or not enough? Was the work itself poorly defined?

When an agent produces a strong result, ask what made it possible. Was the brief unusually clear? Did someone provide a useful example? Was the right source material available? Did the review happen early enough to matter?

Managers improve when they can see the work across assignments, not just the final answer in front of them. They notice recurring rework, repeated questions, expensive detours, and tasks that keep returning for human correction.

That view changes the conversation. Instead of asking whether an individual is good or bad at using AI, we can ask which workflows are improving, which ones are stuck, and what the team should change next.

Someone still has to pay the salary

Agents consume resources. They receive assignments. They produce work. Their output is reviewed, accepted, corrected, or rejected. Their performance depends heavily on the instructions, context, tools, and boundaries they are given.

They may not be employees, but they are becoming part of the workforce. Someone still needs to manage that workforce. And someone has to pay its salary.

That is why the new job cannot be separated from responsibility. We are not only deciding whether to use an agent. We are deciding what work it should do, how much trust it should receive, and what the organization should learn from the result.

The bill is only one signal. The real management question is whether the work was worth assigning, whether the result was useful, and whether the next assignment can be better because of what we learned.

The new management question

Companies may eventually operate with fewer people performing every task manually and more people supervising increasingly capable systems. But those systems will not steer the ship by themselves.

Someone must decide where they should go. Someone must give them the right context, review their work, recognize when they are lost, and decide whether the result was worth the resources consumed.

That responsibility is spreading beyond traditional managers. Developers, designers, salespeople, recruiters, product managers, and many other knowledge workers are taking it on—often without formally recognizing that their role has changed.

Every knowledge worker is becoming an agent manager.

A question worth asking is: Does this AI employee deserve its salary? The answer is not hidden in a usage total. It is found in the quality of the assignment, the usefulness of the result, and whether we are getting better at managing the work between them.