The Shape of Agency

Many organizations say they want their people to have agency. Few have thought carefully about what they mean by it.

At its strongest, agency is the capacity to act — to make decisions that matter, to commit the organization to a course, to spend its money or its reputation, to change what it does without asking permission first. An employee with real agency is someone whose judgment is trusted enough that the organization will live with its consequences. That’s a meaningful thing to give someone, and most companies are more cautious about it than their values pages suggest.

The caution isn’t unreasonable. Agency, at scale, is how organizations move. It’s also how they break.

What agency is, and isn’t

There’s a version of agency that gets confused with autonomy, and another that gets confused with permission. Both miss the point.

Autonomy is being left alone. Permission is being allowed to do a specific thing. Agency is different: it’s the ability to take initiative in a space where the outcome isn’t predetermined, and to be the one whose choices shape what happens next. A finance team that can approve invoices up to a fixed amount has permission. A finance team that can redesign how the company handles vendor relationships has agency.

The distinction matters because organizations routinely confuse the three. They tell people to “act like owners” and then route every meaningful decision through three layers of approval. They celebrate initiative in the abstract and discourage it in the specific. The result is the experience most knowledge workers recognize: the formal scope of one’s role is large, but the actual range of consequential action is small, because the cost of being wrong about anything important has not been bounded in advance.

Why bounded action is the only kind that scales

This is the part that gets missed. Real agency in an organization is not the absence of constraint. It is the presence of well-designed constraint.

Consider the employee who has been told they can spend up to a certain amount without approval, on anything they judge useful. That person has more practical agency than a colleague with no spending limit but a requirement to justify every expense in advance. The first has a bounded space inside which their judgment is trusted absolutely. The second has an unbounded space inside which their judgment is trusted not at all.

The same pattern repeats everywhere agency actually works. A team can ship to production whenever they want, inside an environment where shipping a bad change is recoverable. A salesperson can negotiate terms within a defined range. A manager can hire within a budget. In each case, the constraint isn’t what limits the agency. The constraint is what makes the agency safe to grant, which is what makes it possible to grant at all.

Organizations that try to extend agency without designing the bounds end up doing one of two things. They grant broad latitude and then quietly hope nothing goes wrong — which works until it doesn’t, after which they retract the latitude from everyone. Or they grant the latitude formally and revoke it informally, through the slow accumulation of approval processes that don’t appear on any org chart but reliably prevent anyone from doing anything that matters.

The companies that get this right are the ones that have done the harder work: figuring out where the truly irreversible decisions live, and designing the rest of the space so that being wrong is survivable.

The new actor in the room

This problem of agency design has been around as long as organizations have. What’s changed is that organizations are now extending agency to a new kind of actor, and the old assumptions don’t carry over.

When agency is granted to a person, the design of the bounds can be partly implicit. People deliberate. They ask colleagues. They sleep on things. They notice when something feels wrong and slow down, often without being able to articulate why. Much of what keeps human agency from going badly is not the formal guardrails but the natural latency of human judgment — the gap between being able to do something and actually doing it.

AI agents collapse that gap. An agent acting on behalf of the organization can take a thousand actions in the time a person takes one. And the implicit guardrails that quietly compensated for over-broad human permissions — the fact that people don’t usually think to do most of the destructive things they technically could, the fact that they pause before consequential actions, the fact that someone else tends to notice — do not apply.

This is not a claim about whether agents are reliable. It’s an observation about what agency means when the actor exercising it doesn’t deliberate. Every capability granted to an agent will be used. Every system it can reach will, eventually, be reached. The space of possible actions stops being theoretical and becomes something closer to inevitable. Whatever bounds existed only implicitly, in the slowness and self-doubt of human work, are no longer there.

Consider production database access during an incident. The engineer queries the few rows that answer their question. An agent investigating the same incident may pull the entire customer transactions table — not maliciously, just because retrieving more data is one of the cheapest things it can try. From the perspective of GDPR, PCI, or any other regulatory regime, the engineer’s query and the agent’s bulk pull are the same kind of access. Only the scope is different. And the scope is exactly what the bounds were supposed to govern.

The organizations now piloting AI agents are, in effect, running a stress test on how well they had designed agency for their humans. Where the bounds were real — scoped, enforced, thought through — agents slot into the existing structure and the work expands. Where the bounds were only cultural, or only the accidental product of human hesitation, the cracks show quickly. The agent didn’t create the weakness. It revealed it.

The work ahead

The question stops being “how much can we let this person, or this agent, do” and becomes “what is the space inside which their action is consequential but recoverable, and have we actually built that space.”

For most organizations, that work has been deferred. Human agency was approximate, and the approximation worked. It will not continue to work. The arrival of actors that take every capability they are given and use it at speed will force a reckoning with how agency was ever supposed to function in the first place — which, on reflection, may be the most useful thing about them.

Designing those bounds — for the humans you already trust and the agents you are learning to — is much of what we do at superluminar. If you are working through what accountable agency looks like in this new shape, we’d be glad to think it through with you.