Agents and the autonomy spectrum

Lesson 4 of 5 in What Is AI? From Rules to Reasoning Machines.

The frontier keeps moving. The newest shift — and the one your governance career will be defined by — is from AI that answers to AI that acts. Agentic AI systems decompose a goal into steps, call tools and APIs, hold memory across those steps, and carry them out with limited supervision: booking, buying, coding, negotiating.

Nothing about agents is magic — under the hood sits the same large language model machinery — but delegation changes the governance mathematics. A chatbot's mistake is a wrong sentence; an agent's mistake is a wrong action, possibly the third action in a chain nobody reviewed.

The organising idea is the autonomy spectrum. Place any system on it and its oversight needs follow:

1. Decision support — a human decides; the system informs. (A radiologist reads the scan; the model highlights regions.)
2. Human-in-the-loop — the system proposes; a human approves each output before it takes effect.
3. Supervised autonomy — the system acts; humans monitor and can intervene. (Fraud systems blocking transactions in real time.)
4. Full autonomy — the system acts without review in the moment. (An automated vehicle braking — there is no time to ask.)

Two truths make this spectrum the most practical tool in this module. First, governance intensity should track position on the spectrum — which is why "varying levels of autonomy" sits inside the legal definition of an AI system. Second, position is a design choice that drifts: products launched at level 2 quietly become level 3 as review queues grow and reviewers, subject to automation bias, approve faster than they read. An oversight regime that exists on paper but not in behaviour is the single most common governance failure you will meet in practice.

Key terms: agentic AI, autonomy spectrum, human oversight, automation bias

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.