Skip to content

Orchestration and Delegation

The moment a workflow needs more than one agent, a new problem appears: who decides what work goes to whom, and how does the result get back together correctly?

flowchart LR J["Junior: pass output to input"] --> M["Middle: pick a topology"] M --> S["Senior: isolate context, merge results"] S --> P["Professional: standardize across teams"]

Levels

Level Guide You are done when
Junior Hand off a single output You can wire one step's output into the next step's input with an explicit contract, sequential or parallel.
Middle Pick a topology You can choose between orchestrator-workers, hierarchical, and peer handoff, and justify the cost of adding a sub-agent.
Senior Isolate context, merge results You can scope what each sub-agent sees, and design how conflicting results from parallel workers get resolved.
Professional Standardize delegation across teams You can define shared interface contracts and a deprecation process for sub-agents owned by different teams.

Practice rule

Before adding a sub-agent, write down the specific piece of work it does that the calling step could not do itself in a single step. If the answer is "it's cleaner," that's a code-organization reason, not a delegation reason — a well-organized single step is fine.