Agent Delegation
The manager does not know everything. It should ask the researcher, then ask the developer to implement.
A tech lead who cannot do everything, and should not. They write the ticket, not the patch.
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.Delegation is a manager handing a scoped job to a specialist, then taking the result back. User → Manager → Researcher / Developer / Reviewer — not User → one god-agent → every tool.
- 2.allow_delegation belongs on the manager. Workers stay narrow. This is multi-agent coordination, not a fancier prompt.
- 3.Hierarchical process plus delegation on the manager. Scope the brief: goal, constraints, tools the worker already has.
- 4.Cap rounds so manager and worker cannot ping-pong. Prefer sequential when the handoff order is already known.
Learn elsewhere
- →Processes
- →Multi-agent — Phase 18
Real Example
Scenario
Manager asks research for Stripe API changes, then asks the writer for a migration note. Manager never searches or writes the doc itself.
What you would do
In CrewAI, apply Agent Delegation to this scenario: Manager asks research for Stripe API changes, then asks the writer for a migration note. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Manager delegatesWorkers do not delegateScope the briefCap the ping-pong
Cheat Sheet
Quick recap
quick ref- •Lead assigns
- •Specialists execute
- •No god-agent
- •Log the handoff
Common Mistakes
- ✕Skipping evaluation for Agent Delegation before production
- ✕No logging or tracing around crewai delegation steps
- ✕Ignoring cost and latency implications