Crew + Flow Hybrid
Real products have easy tickets and hard tickets. A crew on every one-line question wastes money. A single agent on a multi-source brief misses review.
ER triage. Nurse (Flow) routes. A specialist (Agent) for a sprain. A trauma team (Crew) for a crash.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Key Takeaways
- 1.The architecture to memorize: Flow provides control, Crew provides autonomous collaboration. User request → Flow classifier → simple path is one Agent, complex path is a Crew.
- 2.Researcher / Writer live inside the complex branch, then Flow reviews and ends. Keep the Flow skinny: classify, route, persist, respond.
- 3.Keep the Crew fat with roles. After the crew, the Flow — not another agent — decides pass/fail, retry, or return.
- 4.That boundary is what you draw in interviews.
Learn elsewhere
- →Flows
- →Build a Research Crew
- →Agent Design Patterns — Phase 17
Real Example
Scenario
FAQ about a term → simple agent. 'Write a 2026 agent-framework comparison' → research + write crew, then Flow saves report.md.
What you would do
In CrewAI, apply Crew + Flow Hybrid to this scenario: FAQ about a term → simple agent. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Flow controls, Crew collaboratesClassify then routeDo not crew a one-linerFlow owns pass/fail
Cheat Sheet
Quick recap
quick ref- •Hybrid is the product
- •Cheap path / expensive path
- •Crew inside a step
- •Flow is the boss
Common Mistakes
- ✕Skipping evaluation for Crew + Flow Hybrid before production
- ✕No logging or tracing around crewai hybrid steps
- ✕Ignoring cost and latency implications