Types of AI agents: choose a pattern for the task
Types of AI agents are useful engineering categories for how a system chooses and coordinates actions. A practical classification separates fixed workflows, tool-using loops, planner-and-executor designs and systems with several agents. These categories describe control and organisation; they do not establish consciousness, general intelligence or a guaranteed level of quality.
Different lists classify agents by different criteria. This guide uses control flow so you can choose an implementation for an actual project. Ask whether the next step is known in advance, must be selected from new information or can be delegated to an independent task.
Key ideas
- Choose a classification axis before comparing categories.
- A fixed workflow is useful when the steps are already known.
- An agentic loop adapts the next action to observations.
- Several agents add coordination duties alongside possible parallel work.
Fixed workflows for repeatable paths
A workflow defines the sequence in application code. For example, read a sample brief, produce a draft, check its required fields and return it for review. A language model may participate in the sequence without choosing its order. This pattern is useful when you can state the transitions before the run begins. Missing input should lead to a defined branch, rather than silently adding new capabilities or turning every exception into an open-ended search.
Tool-using agents for changing information
A tool-using loop lets a model propose the next permitted operation after seeing the current state. A catalogue helper may search by subject and then inspect a returned lesson. The path can change when the first search is empty. Flexibility does not remove boundaries: the tool set, access scope and stopping rule remain defined by the application. Decide how many attempts are appropriate for your exercise and preserve the observations that explain the chosen path.
Planner-and-executor designs
Separating a plan from execution can make longer tasks easier to inspect. The plan describes intended outcomes and dependencies; the executor performs permitted steps; checks report whether each outcome was achieved. A plan should change when evidence contradicts its assumptions. For a lesson route, a proposed advanced topic may need to be replaced after a lookup shows its prerequisite is missing. The plan is a working hypothesis, not permission to run arbitrary actions.
Several agents with distinct responsibilities
A system with several agents may split research into independent questions or separate drafting from review. Define each role's input, output and access. A reviewer needs evidence and acceptance criteria, rather than only another agent's confidence. This pattern is easier to justify when the work can be divided cleanly. If every role repeatedly exchanges the same large context, the added coordination may outweigh the benefit. Test a single-agent baseline before deciding.
A decision checklist
Describe the task without using the word agent. Identify whether the path is fixed, what new information can change it and which outcomes can be checked independently. Choose the smallest pattern that satisfies those requirements. For our teaching catalogue, start with one lookup loop and an explicit missing-record branch. Add planning when there are real dependencies, and add another role when an independent check or separate investigation has a clear job.
- Known sequence: consider a workflow.
- Observation-dependent next action: consider a controlled loop.
- Independent subtasks: compare a multi-agent design with a simpler baseline.
In everyday language
Choosing an agent pattern is like choosing how to organise a kitchen. A recipe gives a fixed sequence. A cook can adapt to the ingredients available. A larger team needs clear jobs and a way to check the finished dish. The team is useful only when the extra coordination serves the meal.
Try it yourself
Classify three teaching tasks: formatting a known record, searching an unfamiliar catalogue and independently reviewing a draft. For each, explain what information changes the next step and which output you can check.
Expected result
Three justified design choices based on task structure, with no claim that a more complex category is automatically better.
Check your answer: Does 'multi-agent' describe intelligence or organisation?
Organisation. It describes several cooperating agent roles, rather than a demonstrated level of intelligence.
Questions
How many types of agents are there?
There is no single universal count because authors classify different properties. A list can describe decision rules, system organisation or capability. State the property you are comparing. For an implementation project, control flow and responsibility often give a more useful starting point than collecting category names.
Are agents the same as general artificial intelligence?
An agent is an execution pattern for choosing actions toward a goal. That does not establish general intelligence or consciousness. A narrow catalogue assistant can use an agent loop while remaining limited to one approved data source and a small set of actions.
Choose the pattern from the task's decision points, dependencies and checks. Complexity should answer a concrete requirement.
Sources and further reading
- Anthropic — Building effective agents ↗Sources checked:
- LangChain — LangGraph overview ↗Sources checked:
- Anthropic — How we built our multi-agent research system ↗Sources checked: