AI GitHub
CrewAI for role-based agent workflows
Assign model-backed roles and tasks within a coordinated workflow.
Key Takeaways
- CrewAI separates crews, which group agents and tasks, from flows that organize application control and state.
- Give each role a clear input and an observable output.
- Use two narrow roles such as researcher and reviewer.
- Extra agents add model calls and can repeat the same mistake if their inputs contain the same error.
How it works
CrewAI separates crews, which group agents and tasks, from flows that organize application control and state. The roles and task outputs define what information is passed between steps.
What you need
Give each role a clear input and an observable output. A task that changes external records needs a separate approval and completion check in the workflow.
What to try first
Use two narrow roles such as researcher and reviewer. Give them a fixture with a known answer and inspect the intermediate results before testing a longer task.
Give the reviewer only the evidence it needs to audit one claim. Track whether passing the entire upstream transcript changes the result or just adds tokens.
Where to look for failures
Extra agents add model calls and can repeat the same mistake if their inputs contain the same error. Compare the workflow against one agent with a clear validation step.
Related Franklin pages
Compare who owns each step and which tools are available at a handoff. A shorter trace with the same validated result can be easier to operate than a larger team of agents.
- Build Python agents that use tools, pass control and expose traces of a run. OpenAI Agents SDK for tools and handoffs
- Build agents with tools and knowledge sources, then organize them into teams or workflows. Agno for agents, teams and workflow services
Browse Franklin for concrete examples of tools and services exposed through MCP. Franklin’s MCP directory
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗