AutoGen — AI Glossary

Reference details

Related terms

• CrewAI
• Agent Swarm / Swarm
• Worker Agent

How it works

1. Agents are modelled as conversable entities that exchange messages, so coordination is expressed as a conversation rather than as a graph.
2. An assistant agent is model-backed; a proxy agent represents a human or executes code, which is how a loop reaches outside the model.
3. Multi-agent conversation is managed by a chat manager that decides who speaks next, so turn order is a policy rather than fixed.
4. Code execution is first-class — an agent can write code and another can run it and return the result, which makes the verification loop native.
5. Termination is a condition on the conversation, so without a well-chosen one the agents will keep talking.

Example: Why termination is the setting that matters

Two agents reviewing each other's work with no termination condition will exchange politely improving messages indefinitely, each turn costing a model call. Ten unnecessary rounds between two agents is twenty calls for no change in the artefact. A condition — a passing test, an explicit completion token, a round cap — is what converts the conversation into a process that ends. The framework will not supply one, and an unbounded multi-agent chat is the characteristic way these systems burn budget.

Commonly confused with

Compared with graph-based orchestration frameworks as though they were variants of one design, and the models differ. A graph framework makes control flow explicit as nodes and edges, so a run is traceable and testable. This makes it emergent from conversation, which is more flexible and harder to reason about — the same trade as workflow versus agent, one level up. It is also not a single-agent framework: the unit here is a group of agents and their speaking policy, which is overhead if the task needs one.

When to use it

Reach for it when:

• Several specialised agents must coordinate and the turn order is better as policy than as a fixed graph.
• Code execution with verification is central, which the framework treats as native.
• A human-in-the-loop participant should sit inside the conversation.

Reach for something else when:

• The control flow is known, where an explicit graph or plain code is testable and cheaper.
• A single agent suffices, making the multi-agent machinery overhead.
• No clear termination condition exists, which lets the conversation run unbounded.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AutoGen — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path