Autonomy is a budget, not a switch

We tend to use the word autonomy as if it were binary: autonomous, or not. Real systems have never worked that way. A thermostat has some autonomy, cruise control has more, an autopilot more still, and a human pilot has broad authority but works inside procedures, training and law.

Agents will be the same. The useful question is not whether an agent is autonomous. It is how much autonomy was delegated, over which actions, for how long, under whose authority, and with what ability to step in.

A ladder, and why it isn't enough

The book uses a simple ladder. At L0 the system answers and a human does everything else. At L1 it assists: it drafts or recommends, and a human acts. At L2 it acts, carrying out a bounded task on its own. At L3 it delegates to other agents. At L4 it organizes ongoing teams of agents. At L5, still speculative, it creates and retires specialist agents as it needs them.

The ladder helps, but a single number hides too much. An agent can be highly autonomous in one way and barely in another. So the book breaks autonomy into five dimensions: how many decisions it makes alone, what actions it may take, how long it may keep going, whether it can deal with other agents, and whether it can change itself, its tools, prompts or policies.

A research agent might make hundreds of analytical decisions while being unable to send an email. An infrastructure controller might make few decisions but hold broad power over production. The second could be more dangerous even though it looks less intelligent. Authority matters more than apparent intelligence.

The autonomy budget

From there an engineering idea follows: give every agent an autonomy budget. It might cap the number of actions, the time, the money, the data it can reach, how many times it may delegate, how many agents it may spawn, its highest privilege, and its blast radius.

Then "is this agent autonomous?" becomes a much more practical question: what budget did we give it? A support agent might get a hundred low-risk actions and no financial authority. A coding agent might get free rein in a sandbox and no production credentials. Autonomy becomes a resource that can be allocated.

"Human in the loop" is three different things

The phrase gets used loosely. A human in the loop must approve before the agent continues. A human on the loop watches and can intervene. A human after the loop reviews the result once the work is done. An irreversible payment may need the first. Routine log analysis may need only the second. A draft report may need only the third.

So the right question isn't whether humans should be involved. It is at which points a human decision should be required. That makes agent design look more and more like access-control design: what can this agent do, who gave it that authority, how long does it last, and who can take it back?

This essay is drawn from Chapter 4 of The Agentic Mind, "The Autonomy Spectrum".