1.2 · Lesson 2 of 7
Multi-Agent Orchestration
What you need to know
Multi-agent orchestration is how several Claude agents share one complex task. The exam is specific about the shape: hub-and-spoke, with a coordinator in the middle.
Hub and spoke
Two roles:
- Coordinator. The hub. It receives the task, splits it, chooses which subagents to call, passes them context, collects results, handles errors, and moves information between them.
- Subagents. The spokes. Each one does a specialised job — web search, document analysis, synthesis, report writing. They take instructions from the coordinator and return results to it.
Every message goes through the coordinator. Subagents do not talk to each other. A shortcut between spokes is the wrong exam answer, including when it looks faster.
That central path gives the exam three properties it asks about:
- Observability. Every message can be logged in one place.
- One error policy. The coordinator applies the same recovery rules to every spoke.
- Controlled context. The coordinator decides what each subagent is allowed to see.
Isolation
A spawned subagent does not receive the coordinator's conversation. It starts with whatever the coordinator writes into its prompt. It does not get:
- The coordinator's system prompt, unless that text is copied in
- Earlier messages from the coordinator's conversation
- Another subagent's results, unless the coordinator passes them
- A shared memory or global store — there isn't one
Invocations are independent of each other too. If the coordinator calls web search twice, the second call does not know what the first one found. Anything a subagent must use has to be in that call's prompt. If synthesis needs the search results, the coordinator pastes them in. Synthesis cannot look them up from a store the agents share.
What the coordinator owns
Four jobs show up on the exam.
- Choose subagents for this query. The coordinator reads the request and picks who to call. It does not run the full pipeline every time. A short factual question may need only web search, not research, analysis, and synthesis. Sending every query through every specialist wastes the turn.
- Split the scope so work does not overlap. Each agent gets a distinct subtopic or source type. One agent can take academic papers while another takes news. Both should not search the same sources.
- Refine in a loop. The coordinator reads the synthesis for gaps. If coverage is thin, it sends search and analysis back out with a narrower question, then runs synthesis again. One pass is not the design.
- Route every message. Spokes report to the hub so logging, errors, and context stay in one place.
The narrow decomposition failure
Learn this pattern by the shape of the miss. A sample exam item asks about a coordinator that splits "impact of AI on creative industries" into visual-arts subtopics only. Music, writing, and film never appear.
The cause is the split, not a downstream agent. Web search was thorough on the topics it was given. Synthesis combined everything it received. Music, writing, and film were never assigned, so they were never researched.
When a report drops entire categories, the scope of the miss points at the coordinator. A gap in breadth is a decomposition bug. A gap in depth inside an assigned topic is a different question.
Worked example: renewable energy
The task is "renewable energy technologies." The coordinator assigns "solar panel efficiency" and "wind turbine design." Each subagent returns careful, sourced work on its assignment.
The report is strong on solar and wind and silent on geothermal, tidal, biomass, and fusion. Search quality and synthesis quality are not the cause. Those topics were never handed out.
The repair is a wider split. Better queries, a stronger synthesis model, and extra subagents still operate on the assignments the coordinator wrote.
Exam traps
Blame the downstream agents when the report misses whole categories
Specialists research what they are assigned. If the coordinator only assigned solar and wind, no specialist can cover geothermal or tidal. Follow the gap back to the split.
Assume a subagent inherits the coordinator's history or shares memory
Each invocation starts with the prompt it was given. A second call does not remember the first. Other agents' results are invisible unless the coordinator pastes them in.
Let subagents message each other to save a hop
On the exam, every message goes through the coordinator. That is what keeps logging, error policy, and context in one place. A direct spoke-to-spoke link is the wrong answer.
Add more subagents to fix a narrow split
Extra specialists still receive the assignments the coordinator writes. If those assignments omit fusion, another agent does not discover fusion. Widen the decomposition.
Practice scenario
A multi-agent research system writes a report on renewable energy technologies that only covers solar and wind. Each subagent produced thorough, well-sourced work on the topic it was given. Web search returned relevant results for every query it received. Synthesis accurately combined everything it was handed. What most likely caused the coverage gap?
Build exercise
Build a hub-and-spoke research coordinator
What you will learn
- Why every message in this architecture goes through the coordinator
- Why isolation means each prompt has to carry its own context
- How a broad split avoids the narrow-decomposition failure
- How a refinement loop finds gaps and sends them back out
- Why a missing category is diagnosed at the coordinator's split
Step 1
Create a coordinator that takes a broad topic
Write a coordinator that accepts a research topic and returns a structured report.
Why: The coordinator is the hub. The exam tests that this agent owns decomposition, subagent choice, and aggregation.
You should see: A function that takes a topic string and returns a report. Its system prompt describes it as the orchestrating hub.
Step 2
Decompose into at least five subtopics
Break the topic into five or more distinct subtopics that cover the full breadth of the subject before any specialist runs.
Why: Narrow decomposition is a named exam failure. A renewable-energy split that is only solar and wind drops whole categories. Incomplete scope traces back to this step.
You should see: A decomposition that yields five or more subtopics. For renewable energy, the list includes solar, wind, geothermal, tidal, biomass, and fusion.
Step 3
Spawn web search and document analysis with explicit prompts
Call two subagents. Put the assigned subtopic, the research goal, and any earlier findings into each prompt.
Why: A subagent starts blank. If its output is thin, check what the coordinator put in the prompt.
You should see: Two invocations. Each prompt contains the subtopic, the goal, and any prior-agent context you intend that agent to use.
Step 4
Aggregate both results and score coverage
Merge the two result sets and compare them to the original subtopic list.
Why: This is where a refinement loop starts. Gaps found here are what get sent back out.
You should see: A combined assessment that marks each subtopic well covered, partial, or missing.
Step 5
Re-delegate until the gaps close
If coverage is incomplete, send targeted follow-ups for the missing subtopics and score again.
Why: A coordinator evaluates and sends work back. A one-shot dispatcher does not. The exam treats that loop as part of the job.
You should see: A loop that finds gaps, queries the subagents for those subtopics, and stops when coverage is enough or a max iteration count is hit.
Step 6
Test renewable energy technologies
Run the coordinator on that topic and check the report for all six energy types.
Why: This is the exam's narrow-decomposition case. Output that only has solar and wind means the coordinator's split was too small.
You should see: Sections on solar, wind, geothermal, tidal, biomass, and fusion, and a coverage check that reports those six as complete.
Sources
- Claude Agent SDK overview — Anthropic
- How we built our multi-agent research system — Anthropic