The short answer
AI agents are classified in two ways. The first is the established academic taxonomy that splits agents by decision mechanism into five types. The second is an operational taxonomy businesses use, splitting agents by how they are triggered and run. Holding both prevents a common confusion: calling a simple system "autonomous" merely because it speaks fluently.
The classic five
1) Simple reflex agent
Acts on the current input alone through condition–action rules. It keeps no internal state and has no notion of what happened a moment ago. Fast and predictable, it fails when the present input is not enough.
2) Model-based reflex agent
Maintains an internal model of the world state, so it can act even when not everything is visible right now. Operationally: an agent that remembers a ticket was already escalated and does not repeat the step.
3) Goal-based agent
Rather than only reacting, it weighs options by how close they bring it to a stated goal and plans a sequence of steps. This is the pattern behind most genuinely useful business agents today.
4) Utility-based agent
When several paths reach the goal, it compares them against a utility function: fastest, cheapest, least risky. In practice that trade-off shows up as explicit priorities you write, such as "prefer the resolution that does not require financial approval".
5) Learning agent
Improves behaviour over time from feedback. In business settings learning is normally bounded and supervised: refreshing knowledge, tuning instructions, and reviewing failed examples — not free-form self-modification in production.
Type table
| Type | Decision basis | Example | Best for |
|---|---|---|---|
| Simple reflex | Rule on current input | Route any message containing "invoice" to finance | Fixed, unambiguous tasks |
| Model-based reflex | Input plus internal state | Not re-escalating a ticket already escalated | Multi-step paths |
| Goal-based | Planning toward an outcome | Qualifying a lead through to a booked meeting | Most business processes |
| Utility-based | Trade-off between options | Picking the cheapest channel with the best response rate | Decisions with alternatives |
| Learning | Feedback loop | Improving screening questions after interview outcomes | High-volume repeat processes |
The practical business taxonomy
- Conversational agent: invoked by a human through a chat surface (support, internal helpdesk, employee assistant).
- Task agent: invoked by an event or request to run one process end to end (screen a CV, issue a quote).
- Scheduled agent: runs on a cadence with no direct request (weekly report, daily sweep of pending orders).
- In-team agent: part of a multi-agent system exchanging outputs with other agents under coordination.
How the two meet
An ecommerce support agent is operationally a conversational agent and, in decision terms, a goal-based agent with reflex rules at the boundary — for example, any refund above a threshold goes to human approval. The mix is deliberate: planning gives flexibility, hard rules give safety.
Choosing the right type
- Describe the task as an end state, not as steps: what must have changed in your systems when it finishes?
- Set the risk level: is the action reversible? If not, add a human approval point.
- Identify the knowledge source: documents, database, or an external system? That determines the tools required.
- Start narrow and measure, then widen — instead of building an agent that "does everything" on day one.
For ready-made examples grouped by role and department, browse the agent marketplace; for the underlying concept, start from the AI agent page.