Types of AI agents: a taxonomy that actually helps

Search types of AI agents and you get the textbook list — simple reflex, model-based, goal-based, utility-based, learning. It’s a fine framework for a syllabus and nearly useless for deciding what to build, because the categories describe internal mechanics rather than anything you choose between. Here’s a taxonomy organised around decisions you actually make.

Why the textbook split doesn’t help

The classic five come from decades-old AI theory and classify agents by how they process input. That was the right question when the hard part was the reasoning engine.

With a capable language model, the reasoning engine is a given. The hard parts moved: how much the agent does without you, how wide its access is, and how far ahead it plans. Those are the axes worth naming, because each is a lever you set deliberately.

By autonomy: who’s in the loop

The most consequential axis. It decides your risk posture more than any capability does.

Assistive. The agent proposes; a human does. Drafts a reply, suggests a fix, surfaces a summary. You keep every decision. Lowest risk, and often the right starting point even for systems that will grow more autonomous.

Supervised. The agent acts, but irreversible actions stop for approval. Reads and reversible writes run freely; sends, deletes and payments wait for a human. This is where most production agents should live — the autonomy is real, the blast radius is bounded.

Autonomous. The agent acts end to end without per-action approval. Appropriate only when every action it can take is reversible or trivially cheap to get wrong. Far rarer in practice than the marketing implies.

The mistake is treating autonomy as a maturity level to climb. It’s a per-action property: the same agent can be autonomous over reads and supervised over sends. Gate on effect, not on confidence.

By scope: how much it can reach

Single-purpose. One job, a handful of tools — a scheduling agent, a triage agent. Predictable, easy to test, easy to trust. Most of the reliable agents in production are this.

Domain. A bounded area with several related jobs — everything email, or everything in your codebase. More useful, more surface to get wrong.

General. Access across many domains and hundreds of tools, choosing what’s relevant per request. The most powerful and the hardest, because tool selection itself becomes a retrieval problem once the set is large.

Scope trades capability against reliability directly. Wider agents do more and fail in more ways, which is the tension the next axis exists to manage.

By planning horizon: how far ahead it commits

Reactive. One step at a time — observe, act, observe again. Robust, because there’s no long plan to fall apart, and it self-corrects when the world changes mid-task. The right default.

Planning. Lays out a multi-step plan, then executes it. More efficient when steps are genuinely independent and parallelisable, more fragile when reality diverges from the plan and the agent keeps following a map that no longer matches the ground.

The instinct that more planning is more sophisticated is often wrong. A reactive agent that checks after each step usually beats a planner that commits to six steps and discovers at step four that step two failed.

Single vs many

Orthogonal to all of the above: one agent or several.

A single agent covers more than people assume. Multiple agents earn their coordination cost only under specific conditions — parallelism, real specialisation, or credential isolation. “Multi-agent” is an implementation choice, not a category of capability, and reaching for it first is the most common over-engineering in this space.

Putting the axes together

A real system is a point in this space, not a textbook label. A production triage agent is supervised, single-purpose, reactive, single. A research system might be supervised, domain, planning, multi. Naming all four axes tells you more than any single category name ever could.

Agent type	Autonomy	Scope	Planning horizon	Example
Drafting assistant	Assistive	Single-purpose	Reactive	Suggests replies; a human sends every one
Triage agent	Supervised	Single-purpose	Reactive	Labels and archives mail; unsubscribes wait for approval
Scheduling agent	Supervised	Single-purpose	Planning	Proposes a slot across calendars; the invite waits for a tap
Domain agent	Supervised	Domain	Reactive	Everything email: triage, drafts, follow-ups, one tool set
Research agent	Supervised	Domain	Planning	Fans out across sources, synthesises; reads only
General agent	Supervised	General	Reactive	Picks tools per request from hundreds; gates on effect
Background automation	Autonomous	Single-purpose	Reactive	Reversible-only actions: labelling, logging, filing

The practical value is that each axis is a dial you set on purpose. Start conservative — assistive or supervised, single-purpose, reactive, single — and widen one axis at a time, only where you’ve watched the current setting strain. Widening several at once is how you get a system nobody can debug.

Where marketing muddies it

Two labels worth translating.

“Autonomous agent” usually means supervised-with-good-defaults, not truly unsupervised. That’s a feature, not a shortfall — genuinely unsupervised action over irreversible operations is a liability, not a selling point.

“AI employee” is a scope-and-autonomy claim dressed as a metaphor. Ask which of the axes above it actually occupies. A capable agent is closer to a well-scoped tool with judgement than to a person, and the metaphor mostly obscures the questions you should be asking.

The useful frame stays the same: how much does it do alone, how much can it reach, and how far ahead does it commit? Answer those and you’ve described the agent precisely — and specified most of what you need to build it.

See also: what an AI agent is and how agents differ from workflow automation.