Artificial intelligence has long promised to augment human ability. Promises are cheap. The practical change is that some systems now do more than answer: they receive a goal, form a plan, use tools, and revise their approach when the plan meets reality, where plans have traditionally gone to acquire character.

What Makes an Agent?

Teams use the word agent loosely, but practical architectures often combine four abilities, even when researchers disagree about where one ends and another begins (Yao et al., 2023; Park et al., 2023):

  • Perception: receiving information from an environment
  • Reasoning: choosing steps toward a goal
  • Action: changing something outside the chat window
  • Memory: carrying useful context across steps or sessions

Together, these abilities move software into the workflow. An agent may search the web, run code, update a record, or schedule a meeting in pursuit of a goal stated in plain language. The distinction matters because an answer can be wrong on a screen. An action can be wrong in the world.

From Tasks to Loops

Automation has always divided work between people and machines. Labor economists often describe jobs as bundles of tasks because technology may replace some activities while making others more valuable (Autor, 2015). Agentic systems add a complication: they can enter the connections among tasks.

Consider a customer-support case. A conventional assistant might summarize the account or draft a reply. A support agent could read the history, decide which policy applies, issue a refund, update the customer record, and schedule a follow-up. None of these steps is especially dramatic. Their combination is.

The unit of delegation has widened from write this response to resolve this case.

That gives the system latitude over sequence, method, and exception handling. It also makes the work harder to inspect. A person may see the final resolution without seeing the discarded evidence, the policy interpretation, or the moment the agent decided that an unusual case was ordinary enough to continue.

This is where agentic AI differs from a faster tool. The system begins to participate in how the work is organized.

Work Includes Noticing

A job is more than its completed outputs. Work is also where people encounter edge cases, acquire judgment, learn which rules bend, and discover that the official process has only a passing acquaintance with the real one.

When an agent handles the ordinary sequence, a person may gain time for harder work. They may also inherit an uninterrupted queue of exceptions chosen by a system whose earlier decisions they did not observe. The first arrangement can increase a worker’s reach. The second can make expertise harder to develop and exhaustion easier to measure.

Neither outcome follows automatically from the technology. It depends on who chooses the boundary, whether workers can inspect the agent’s reasoning, and whether efficiency gains become time, autonomy, surveillance, or simply more work.

Delegation therefore changes the distribution of knowledge. The person who once performed the process may become the person who approves its residue. Over time, an organization can lose practical understanding even while its dashboard reports that every case closed faster.

Which Decisions Require Human Authority?

The usual question asks what work will remain human. That makes human work sound like whatever the machine has not learned to do yet. A better question concerns authority: which decisions should people continue to understand, contest, and own even when software can execute them?

Direction determines the goal. A system can pursue a measurable target without knowing whether the target deserves pursuit, what the measure conceals, or who was absent when it was chosen.

Judgment connects a rule to a situation. It recognizes when the available categories fit badly, when an exception contains important information, and when a successful metric has become detached from its purpose.

Values determine which tradeoffs are acceptable and whose interests count. They enter through objectives, policies, data, escalation rules, and the institutional decision to automate one person’s work rather than another’s.

These capacities are not possessed by an abstract entity called the human. Workers, managers, customers, builders, and affected communities may want different things. A system can feel empowering to the person directing it and coercive to the person being evaluated by it. Human-centered design begins by asking which human occupies the center and who has been placed just outside the diagram.

Accountability Can Stay Behind

Authority and responsibility do not always move together. Madeleine Clare Elish describes moral crumple zones in which a person absorbs blame for the failure of an automated system despite having little meaningful control over it (Elish, 2019). Agents can widen this gap because their actions extend across more of the workflow.

A worker may be expected to approve hundreds of agent-generated decisions without enough time or context to examine them. A manager may sign off on a process whose exceptions are invisible. A customer may receive a consequential decision and find nobody able to explain which rule produced it.

All three arrangements contain a human in the loop. Presence alone does not create control.

Meaningful control requires time to inspect, authority to intervene, and enough understanding to challenge the system’s choice. It also requires an organization willing to treat disagreement as information rather than delay.

Competence Does Not Confer Authority

AI safety research has documented related failures under names such as reward hacking and harmful side effects: a system satisfies the stated objective while violating the purpose that made the objective worth pursuing (Amodei et al., 2016). Greater capability can intensify this problem. A competent system reaches the wrong destination more efficiently.

Reserving every consequential action for direct human execution would discard much of what makes agents useful. The design task is to connect autonomy to evidence and consequence.

Systems can begin with narrow authority and earn more through reliable performance. People need to see what an agent did, why it did it, and what changed. Decisions that are costly, irreversible, or morally significant deserve deliberate interruption. Failure must remain recoverable where recovery is possible.

Most of all, the people affected by a workflow need ways to question it. An appeal process that returns the decision to the same model with a longer prompt is an impressive loop. It is not an independent review.

Where the Leverage Goes

Agentic systems may expand what one person can accomplish. They may let a small team investigate more possibilities, remove tedious coordination, and make specialized knowledge easier to use. Those benefits are real possibilities. They are not evenly guaranteed.

The same system can centralize decisions, standardize work around one institution’s assumptions, or give managers detailed visibility into workers while giving workers little visibility into the agent. Whether automation complements expertise or erodes it is partly a technical question. It is also a question of ownership, bargaining power, and who is allowed to refuse the arrangement.

Productivity tells us how much output a system produces. It does not tell us who gained time, who lost discretion, who learned less, or who became easier to replace.

Those are properties of the work too.

What to Watch

The most useful evidence about agents will come from ordinary deployment, after the demonstration and before the institutional story has settled. We should ask:

  • Which decisions did the agent make rather than execute?
  • What can workers and affected people inspect, challenge, and reverse?
  • Who gains time, knowledge, discretion, and bargaining power?
  • Who carries the risk when the system fails?
  • What did the demonstration quietly omit?

The rise of agentic AI is a change in capability. Its meaning for human work will be decided through job design, product choices, institutional incentives, and conflicts over authority. The technology supplies new ways to delegate. People still decide what delegation is for, though some will have considerably more influence over that decision than others.

The agentic age will not keep humans at its center through good intentions alone. Centers are designed. So are exits.

Sources