what are the 7 types of ai agents?

where the numbers come from

The five-type taxonomy is from Russell and Norvig's Artificial Intelligence: A Modern Approach, the standard university textbook, and it predates language models by decades. It classifies agents by how much internal state and reasoning they carry, not by what they are made of.

The seven-type version is a web-era extension. It keeps the five and bolts on two organisational patterns. That is why the count varies between four, five, six, and seven across articles: everyone is extending the same base list by a different amount.

the seven, in order of increasing sophistication

1. Simple reflex agent. Maps the current input straight to an action with condition-action rules. No memory. A thermostat. Breaks the moment the right action depends on what happened earlier.
2. Model-based reflex agent. Keeps an internal model of the world so it can act on parts it cannot currently see. A robot vacuum that remembers which rooms it has covered.
3. Goal-based agent. Has an explicit goal and searches for a sequence of actions that reaches it. A route planner. Slower than reflex, far more flexible.
4. Utility-based agent. Has a utility function so it can rank several goal-satisfying plans and pick the best trade-off. A pricing engine balancing margin against conversion.
5. Learning agent. Improves its own behaviour from feedback. Split into a performance element that acts and a critic that scores.
6. Hierarchical agent. A layered structure: high-level agents decompose goals and hand sub-goals to lower ones. This is a topology, not a new kind of reasoning.
7. Multi-agent system. Several agents, each with its own goals, coordinating or competing. Also a topology.

why the taxonomy feels off for llm agents

A modern language-model agent is not any single row. It is usually goal-based in structure, utility-flavoured in how it picks between plans, model-based because it carries memory, and not really a learning agent at all, because the weights do not change between your runs. What people call learning is memory and retrieval.

So the taxonomy is still useful for one thing: telling you which capability a system actually has. If a vendor calls something a learning agent, ask what updates. If the answer is a notes file, that is memory, and that is fine, but it is not the textbook meaning.

the two that matter commercially

Rows 6 and 7 are where the money is right now, because they describe how to get past the ceiling of a single agent holding fifty tools. Hierarchy gives you delegation. Multi-agent gives you specialisation and a second opinion. Most production agent work in 2026 is one of those two shapes wearing a product name.

how this works in aldena

Aldena is rows 6 and 7 made concrete. Agents are goal-based with durable memory, arranged in an org chart so a manager decomposes work and delegates it down its own line, which is the hierarchical pattern. Several specialists then work the same project inside one room, which is the multi-agent pattern.

There are eleven prebuilt roles, and a room on the starter plan holds up to ten agents. What each one can touch is set by tool permissions, so the utility trade-off you care about most, speed against blast radius, is a setting rather than a hope.