Notes
🔍

Autonomy (Agent)

Autonomy is easy to picture as a switch — either the software asks you first or it doesn't. In practice it is a dial, and where you set it is one of the central design decisions in building an agent. Autonomy is the extent to which a system chooses its own next action rather than executing one you specified.

What the dial actually controls
Several independent things are bundled under the one word:

  • Decision scope — may it choose which tool to call, or only fill in the arguments for a call you fixed?
  • Horizon — how many steps may it take before coming back to you?
  • Irreversibility — which actions may it perform unsupervised, and which need an approval gate?
  • Resources — how much time, money and how many tokens may it consume unasked?

A workflow with a fixed sequence of model calls has almost none of this; an agent handed a repository and a week has a great deal.

Why the setting is a trade
Autonomy is what makes an agent worth having: nobody gains from a system that needs confirmation for every file it reads. It is also the entire source of agent-specific risk — a loop with nobody in it can burn a budget on a failing retry, obey an instruction hidden in a page it fetched, or act confidently on a misread state. Permission scoping, step limits, sandboxes and human-in-the-loop review all exist to buy back control without turning the agent back into a form you have to fill in.

Autonomy is the degree to which an agent decides its own next action instead of executing a step a human specified — spanning which tool to call, how many steps to take, and which actions it may perform without approval. It is a dial, not a property: raising it increases what the agent can finish unattended and simultaneously widens what it can get wrong unnoticed.

Think about handing a job to a new assistant. On the first day you check everything. Later you might say "book the whole trip" and only hear back when it's done. What changed is how much you let them decide for themselves.

Autonomy is that same dial for software. At the low setting it asks before every move. At the high setting it picks its own next step, uses the tools it needs, and keeps going until the job is done.

Turning the dial up gets more done without you — and means mistakes can happen with nobody watching, which is why such systems are given spending limits and a list of things they must ask about first.