Multi-Agent Orchestration — Detail & Examples

Part of Agentic Architecture · overview: domain-1.html

In this page

A full coordinator/subagent handoff example with a sample return shape, a use-case comparison of the three topologies, and the most common handoff bug to avoid.

1.3aA hub-and-spoke handoff, end to end

The most common production pattern. A coordinator owns the goal; each subagent gets one narrow, self-contained job and its own isolated context — it never sees the coordinator's full transcript, and the coordinator never sees the subagent's full tool-call history, only its summary.

WORKED EXAMPLE · COORDINATOR / SUBAGENT HANDOFF

COORDINATOR TASK"Produce a market-entry brief for launching in Germany."

Coordinator decomposes into 3 independent research questions and spawns one subagent per question: regulatory requirements, competitor landscape, localization needs.
Each subagent works in total isolation — the regulatory subagent has no idea the competitor subagent exists, and doesn't need to.
Each subagent returns a condensed summary, not its raw research transcript — typically a few hundred words, not the dozens of tool calls it took to get there.
Coordinator synthesizes the 3 summaries into one brief, resolving any contradictions between them itself.

Subagent return shape:{ finding: "...", confidence: "high", sources: [...], open_questions: [...] } — structured, not a raw transcript.

1.3bChoosing a topology

Three shapes, three use cases

Hub-and-spoke (example above): independent subtasks reporting to one coordinator. Easiest to reason about and debug — use this by default.
Pipeline: a document translation and localization task where each stage's output is literally the next stage's input — draft → translate → culturally adapt → legal review. Order is fixed and each stage transforms, not just reports.
Peer-to-peer: two negotiating agents (e.g., a buyer agent and seller agent exploring a deal) that need to talk directly. Most flexible, but the least predictable — harder to log, harder to bound, and easy to get an unproductive back-and-forth loop. Reserve for cases hub-and-spoke genuinely can't express.

1.3cThe handoff bug to watch for

The most common orchestration mistake isn't choosing the wrong topology — it's a subagent returning too much. A subagent that dumps its entire 40-call tool transcript back to the coordinator has technically "reported back," but it has also defeated the point of isolating it: the coordinator's context now balloons with detail it doesn't need, which is exactly the context-degradation problem covered in Domain 5.1.

← 1.2 Task Decomposition Agentic Architecture Overview 1.4 Reliability & Oversight →

Claude Certified Architect — Study NotesDWG 1.3