CrewAI — AI Glossary

Reference details

Related terms

• AutoGen
• Worker Agent
• DAG (Directed Acyclic Graph)

How it works

1. Agents are declared with a role, a goal and a backstory, which together become the system prompt that conditions that agent's behaviour. The role framing is the framework's central abstraction.
2. Tasks are declared separately, each with a description, an expected output and an assigned agent, so the work and the worker are specified independently.
3. A crew composes agents and tasks with an execution strategy: sequential, where each task's output feeds the next, or hierarchical, where a manager agent assigns work dynamically.
4. Each agent carries its own tools, so capability is scoped per role rather than exposed to every agent in the crew.
5. Delegation is an action an agent can take: asking another agent to handle something, which adds a model turn and a context transfer at each handoff.
6. The structure is declarative, which makes a multi-agent system quick to express and means the per-handoff cost is easy to incur without noticing — role prompts and transfers are carried on every request they touch.

Example: What the role abstraction costs per request

A crew of 5 agents, each with a role, goal and backstory totalling about 300 tokens, carries 5 x 300 = 1,500 tokens of role definition across the crew. That is modest in itself. The cost arrives with delegation: each handoff is an additional model turn, so a run with 3 handoffs adds 3 model turns, and at 800ms per turn that is 3 x 800 = 2,400ms of latency spent on coordination rather than work. Against a single-agent run of the same task at, say, 4 turns totalling 3,200ms, the crew version is 4 + 3 = 7 turns at 5,600ms — a 5,600 / 3,200 = 1.75x increase for the same output unless the specialisation actually improves it. The question that decides whether a crew is worth it is therefore whether role separation produces a better result than one agent with all the tools, and that is measurable: run both on the same tasks and compare. The default assumption should be that it does not, because a single agent with the full context has information the delegating crew deliberately divided, and specialisation has to earn back the 75%.

Commonly confused with

Role prompting and genuine specialisation are not the same thing, and the role abstraction makes it easy to assume the first delivers the second. Giving an agent the backstory of a senior financial analyst changes its tone and its framing; it does not give it financial data, a calculator, or training it did not have. Real specialisation comes from different tools, different context and sometimes different models — which the framework supports and which the role fields do not supply. A crew whose agents differ only in their backstories is one model talking to itself in several registers, at several times the cost.

When to use it

Reach for it when:

• Workflows that genuinely decompose into roles with different tools and different permission scopes.
• Rapid prototyping of a multi-agent structure, where the declarative form is faster than writing the orchestration.
• Sequential pipelines with clear handoff points and a defined output per stage.
• After measuring a single-agent baseline, so the coordination cost is being paid for a demonstrated gain.

Reach for something else when:

• Tasks a single agent with the full tool set handles, where delegation adds turns and divides context for nothing.
• Agents differing only in their role text, which is one model in several registers rather than specialisation.
• Latency-sensitive paths, where each handoff adds a model turn to the critical path.
• Where the handoff would lose context the next stage needs, which is the characteristic failure of divided-context designs.

Continue your research

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

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