Autonomy is a delegation problem

The framing has become familiar enough to feel settled. We have moved, the story goes, from chat to action — from models that wait for a prompt, answer, and fall silent, to agents that plan, call tools, and execute until a goal is met. The advisor becomes the assistant who actually does the thing. By that telling, 2026 is the year autonomy stops being a demo and becomes the default, and the telling is basically right. Systems that act instead of merely answering are arriving, they work well enough to matter, and the diffuse, reversible middle of most organizations’ work will be better for them.

But the standard version treats this as a capability upgrade — more done, faster, with less human friction — and that framing quietly hides the thing that actually changed. The moment a system acts instead of advising, it is making decisions and carrying them out. You have not bought a faster worker. You have delegated authority. Almost everything that is hard about agents, and almost everything the excitement skips, lives inside that sentence.

Advice you can ignore; action you can’t

A language model produces signal. It drafts, suggests, predicts — and you remain the decision-maker, free to take the output or throw it away. The worst case of a bad suggestion is a wasted minute and a discarded paragraph. An agent removes you from that step by design. It does not suggest the refund; it issues it. It does not recommend the email; it sends it. It does not flag the record; it changes it. The same underlying mistake that used to be a bad draft you caught is now an executed consequence you discover later.

This is the real content of “autonomy.” It is less an increase in capability than a relocation of the decision — out of a person’s hands and into a process that most organizations have not instrumented to ask whether the decision was good. The technology didn’t just get more able. It moved the point of commitment to a place where nobody is looking.

The question was never capability

Most of the agent conversation fixates on can it — can it plan, use the tools, recover when a step fails, coordinate with other agents. Those questions are now answered well enough that they are no longer the bottleneck. The unanswered question is a different one: should this decision be delegated, and how would you know the delegation is working? Delegating a decision to a system you cannot evaluate is delegating it blind, and autonomy is dangerous precisely because blind delegation feels like progress. The dashboard fills with completed tasks, throughput rises, the line goes up — and none of it tells you whether the judgments the agent made were the ones you would have wanted. Task completion is a proxy. Whether the right calls got made is the outcome, and the two come apart exactly where it matters most.

Where autonomy belongs — and where it doesn’t

Delegation is not a maturity level you ascend toward, with full autonomy waiting at the top. It is a property you grant one decision at a time, and the right amount of it is wildly uneven. The decisions safe to hand over are the high-volume, low-consequence, easily reversed ones — the diffuse tail of the work, where a wrong call is cheap to detect and cheaper to undo. Hand those over aggressively; the productivity case is real and the downside is bounded. The decisions that must not be handed over — or only under tight human accountability — are the consequential ones: high-stakes, made under genuine uncertainty, expensive or impossible to reverse.

The hype pushes the opposite. It treats autonomous workflows as a standard to roll out everywhere, and measures sophistication by how much of the human has been removed. That is not sophistication. Granting authority everywhere, including over the decisions you cannot afford to get wrong, is not maturity — it is abdication wearing the costume of progress.

Bounding the authority

If an agent is a delegation, then deploying one responsibly looks like delegating responsibly to a person, only more explicit. Three things hold it together.

Name the authority. State, in advance and in writing, what the agent may decide and act on by itself, and the boundary past which it must stop and ask. An agent without a defined edge to its authority is not autonomous; it is unbounded, which is a different and worse thing.

Instrument the decision, not the task. Completion tells you the agent did something. It does not tell you whether what it did was right. The measurement has to reach the decisions themselves — were they the calls a good human would have made, judged against that baseline — because that is the only signal that separates an agent that works from one that is merely busy.

Keep the revocation path live. Authority you cannot withdraw the instant the evidence turns is not authority you granted; it is exposure you accepted. The willingness and the mechanism to pull the delegation back — fast, without a quarter of change management — is the real price of handing it over. If revoking is hard, you did not deploy an agent. You acquired a dependency.

The loop is not scaffolding

The standard story treats the human in the loop as a temporary inefficiency — present “for now,” “for high-stakes,” destined to fall away as the agents mature. For the reversible middle, that is true, and good. For consequential decisions it has it backwards. The accountable human is not scaffolding that comes down once the building stands. An agent can make a decision; it cannot be accountable for one. Accountability does not delegate. When a consequential automated decision goes wrong, “the system did it” is not a sentence a serious organization can say to a regulator, a customer, or a board. So the human stays — not because the technology is immature, but because the consequences are real, and they will still be real when the technology is flawless.

The quiet part

2026 may well be the year action becomes ordinary, and for most of the work that is simply good news; the case for delegating the cheap, reversible, high-volume decisions is strong and getting stronger. But the framing the revolution arrives wrapped in — autonomy as the removal of the human, a dashboard of completed tasks as proof it worked — is the part that will cost organizations the most, and it will cost them quietly, in decisions that executed before anyone asked whether they were right.

An agent is a decision you have stopped watching. Sometimes that is exactly what it should be. The discipline of this era is knowing precisely which decisions those are — and keeping the nerve to watch the ones that matter, no matter how capably the system offers to take them off your hands.