Hub-and-spoke orchestration

Definition

A multi-agent pattern with one orchestrator ("hub") that spawns and coordinates specialist subagents ("spokes"), each working in isolated context.

Key points

• One orchestrator + specialist subagents; the orchestrator delegates via the Task and Agent tool.
• Fits parallel / isolated-context / independent workstreams (see Subagent decision criteria); each spoke's isolated context prevents Context rot and enables parallelism.
• Parent→child channel is only subagent_type + a single prompt string; children return results but do not share the parent's history.
• For orchestrating dozens–hundreds of agents, use the Workflow tool.
• Decomposition into spokes can be static (prompt chaining) or dynamic (agent decides) — see Task decomposition (chaining vs adaptive).

Why it matters for the exam

• The Multi-Agent Research System scenario is essentially hub-and-spoke; expect questions on when it beats a single agent and how the orchestrator coordinates spokes.

Common gotchas

• Not for sequential/shared-state/single-file work — that's where a single agent wins.
• Over-spawning spokes (esp. Opus 4.6) degrades results; give the orchestrator explicit delegation guidance.

See also

• Subagent decision criteria
• Task and Agent tool
• AgentDefinition
• Task decomposition (chaining vs adaptive)
• Workflow tool

Sources

• netlight-study-guide
• docs-tools-agents