Why the classic taxonomy still matters
Every AI textbook starts with the same classification of types of agents in ai: simple reflex, model-based reflex, goal-based, utility-based, and learning. It feels academic — until you start building production systems and realise the taxonomy maps almost exactly onto the architectural choices you have to make.
The five types and where they fit
- Simple reflex agents — if-this-then-that systems. Use these where the action is purely a function of the current input. Most rules engines are this. Don't over-engineer.
- Model-based reflex agents — same as above but with state. Useful when the right action depends on what happened earlier in the session.
- Goal-based agents — given a desired outcome, plan a sequence of actions. Where most agentic LLM systems sit today.
- Utility-based agents — optimise a continuous reward, not a binary goal. Used in recommendation, bidding, and trading systems.
- Learning agents — improve from outcomes. The endpoint of any closed-loop AI system.
Where conversational agents fit
Conversational AI agents like the ones we build on TalkTaro are typically goal-based with utility scoring on top: they have a goal (resolve the customer's intent), they choose actions (ask, retrieve, escalate), and they score those actions on customer satisfaction signals.