Types of AI agents
There are two ways to sort AI agents. Textbooks rank them by how they think: simple reflex, model-based, goal-based, utility-based, and learning agents. Builders use a simpler split: workflows, where a developer fixes the steps, and agents, where the AI model decides the steps itself.
People group AI agents in two ways — the textbook way and the way builders use when deciding what to make. You only need to recognise the first. The second is the one you'll actually use.
Textbooks sort agents by how much thinking sits between seeing and acting [1]: simple reflex agents follow fixed if-this-then-that rules (a thermostat); model-based agents remember things they can't see right now; goal-based agents plan toward a clear goal; utility-based agents score the possible outcomes and pick the best; and learning agents get better with experience. Think of it as a ladder of growing independence. Most modern AI systems sit near the top.
Workflow or agent: which split matters?
For builders, the first question is workflow versus agent [2]. In a workflow, the developer fixes the path and the AI model fills in each step: chaining (one answer feeds the next step), routing (sort the request, then send it to a specialist), parallelization (run steps side by side, then combine the results), orchestrator–workers (one model splits the task, others do the parts), and evaluator–optimizer (one model writes, another checks, and they repeat). A true agent goes further: the model plans, uses tools, checks the results, and repeats until it decides the job is done.
What matters isn't what the agent is — it's who controls the loop.
Where do AI agents actually work?
You can also sort agents by where they do their work. Coding agents work inside a code project, using the terminal, editors, and tests. Computer-use and browser agents use real screens like a person would — clicking, typing, reading. Voice and support agents hold conversations within set rules. Multi-agent systems put several of these behind a coordinator that splits up and hands out the work [3]. Where an agent works decides everything hard about it: the tools it needs, the ways it fails, and — as the evaluation guide explains — how you have to test it.
Sources
- Russell & Norvig — Artificial Intelligence: A Modern Approach (five agent classes)
- Anthropic — Building effective agents, 19 Dec 2024 (workflow patterns vs agents)
- AWS — What are AI agents? (types of AI agents section)
Frequently asked questions
What are the five classic types of AI agents?
Textbooks list five: simple reflex agents (fixed rules), model-based agents (they remember what they can't see), goal-based agents (they plan toward a goal), utility-based agents (they score options and pick the best), and learning agents (they improve with experience). Most modern AI systems sit near the top of this ladder.
What is the difference between a workflow and an agent?
In a workflow, the developer fixes the path and the AI fills in the steps — chaining them, routing requests, or running steps side by side. In an agent, the model itself plans, uses tools, checks the results, and repeats until it decides it's done. The key question is who controls the loop.
What kinds of AI agents exist in practice today?
Sorted by where they work: coding agents work in code projects through terminals and tests; computer-use and browser agents click and type on real screens; voice and support agents hold conversations within set rules; and multi-agent systems put several of these behind a coordinator that splits up the work.