1.2 Multi-Agent Orchestration

What You Need to Know

Multi-agent orchestration is how you get several Claude agents working together on one complex task. The exam isn’t loose about the shape this takes. It tests one pattern: hub-and-spoke, with a coordinator at the centre.

Hub-and-Spoke Architecture

The architecture has two roles:

Coordinator agent: sits at the centre. Receives the initial task, decomposes it, decides which subagents to invoke, passes context to them, aggregates their results, handles errors, and routes information between them.
Subagents: the spokes. Each one handles a specialised task (web search, document analysis, synthesis, report generation). They receive instructions from the coordinator and return results to it.

The cardinal rule: ALL communication flows through the coordinator. Subagents never communicate directly with each other. Never. Not for efficiency, not for convenience, not for any reason. Every piece of information that moves between subagents passes through the coordinator.

This centralisation provides three things the exam cares about:

Observability — you can log and monitor every message in one place.
Consistent error handling — the coordinator applies uniform error recovery policies.
Controlled information flow — the coordinator decides what context each subagent receives.

The Critical Isolation Principle

This is the single most misunderstood idea in multi-agent systems, and the exam leans on that confusion hard.

Subagents do NOT automatically inherit the coordinator’s conversation history. When the coordinator spawns a subagent, that subagent starts with only what the coordinator explicitly includes in its prompt. It has no access to:

The coordinator’s system prompt (unless explicitly included)
Previous messages in the coordinator’s conversation
Results from other subagents (unless the coordinator passes them)
Any “shared memory” or global state

Subagents do NOT share memory between invocations. If the coordinator calls the web search subagent twice, the second invocation has no knowledge of the first. Every invocation is independent.

So the coordinator has to be deliberate about context. Every piece of information a subagent needs goes in its prompt, explicitly. If the synthesis agent needs web search results, the coordinator passes those results — the synthesis agent can’t “look them up” from a shared store. There’s no shared store.

Coordinator Responsibilities

The coordinator has four key responsibilities that the exam tests:

1. Dynamic subagent selection. The coordinator analyses query requirements and dynamically selects which subagents to invoke. It does NOT always route through the full pipeline. A simple factual question might only need the web search subagent, not the full research-analysis-synthesis chain. Routing every query through every subagent wastes time and resources.

2. Research scope partitioning. When delegating to multiple subagents, the coordinator partitions the research scope to minimise duplication. It assigns distinct subtopics or source types to each agent. For example, one agent searches academic papers while another searches news articles — they do not both search the same sources.

3. Iterative refinement loops. The coordinator evaluates synthesis output for gaps. If the synthesis is incomplete, it re-delegates to search and analysis subagents with targeted queries. It re-invokes synthesis until coverage is sufficient. This is not a single-shot process — it is an iterative cycle.

4. Centralised communication routing. All subagent communication routes through the coordinator for observability, consistent error handling, and controlled information flow.

The Narrow Decomposition Failure

This is a specific exam pattern you must recognise. The exam guide’s sample set includes a question (Question 7) where a coordinator decomposes “impact of AI on creative industries” into only visual arts subtopics, missing music, writing, and film entirely.

The root cause is the coordinator’s task decomposition, not any downstream agent. The web search agent searched thoroughly for what it was assigned. The synthesis agent synthesised everything it received. But the coordinator only assigned visual arts topics, so music, writing, and film were never researched.

The exam expects you to trace failures to their origin. When a multi-agent system produces a report that misses entire categories, do not blame the subagents — check the coordinator’s decomposition.

This pattern applies broadly: if the output is incomplete in scope (not depth), the coordinator’s decomposition is almost always the root cause.

Practical Example: Research System Coverage Gap

A multi-agent research system is tasked with “renewable energy technologies.” The coordinator decomposes this into “solar panel efficiency” and “wind turbine design.” Each subagent produces thorough, well-sourced research on its assigned topic.

The final report is comprehensive on solar and wind but says nothing about geothermal, tidal, biomass, or nuclear fusion. The coverage gap is not because the search was poor or the synthesis was weak — it is because the coordinator never assigned those subtopics.

The fix is not better search queries, not a more capable synthesis agent, and not more subagents. The fix is better coordinator decomposition that covers the full breadth of the topic.

Exam Traps

Practice Scenario

A multi-agent research system produces a report on ‘renewable energy technologies’ that only covers solar and wind power. Each subagent produced thorough, well-sourced coverage of its assigned topic. The web search subagent returned relevant results for every query it received. The synthesis subagent accurately combined all research it was given. What is the most likely root cause of the coverage gap?

A. The web search subagent used queries that were too narrow, so geothermal, tidal, biomass and fusion sources never appeared anywhere in its results
B. The synthesis subagent failed to identify gaps in the research it received and request additional coverage of the missing technology categories from the coordinator
C. The coordinator decomposed the topic into only solar and wind subtopics, never assigning geothermal, tidal, biomass, or fusion to any subagent
D. The document analysis subagent had no access to sources covering the other renewable energy categories, so none of those sections were ever written

Answer & explanation

Correct: C

A — The web search subagent researched exactly what it was assigned and returned relevant results. The issue is not how it searched — it is what it was asked to search for.
B — The synthesis subagent works with the research it receives. It cannot synthesise topics that were never researched. Gap identification is the coordinator responsibility during iterative refinement.
C — The coordinator is responsible for task decomposition. If it only assigns solar and wind as subtopics, no downstream agent can cover the missing categories. The root cause is the coordinator decomposition, not any subagent performance.
D — Source availability is not the issue. The coordinator never asked any agent to research these other energy types. Even with perfect source access, unassigned topics would remain uncovered.

Sources

Claude Agent SDK Overview — Anthropic
Building with Claude API, including the Multi-Agent Research System scenario (Skilljar) — Anthropic

Exam Simulator

Five exam-style multiple-choice questions on Multi-Agent Orchestration. Pick an answer, then open the explanation.

Question 1

A multi-agent research system produces a report on “renewable energy technologies” that only covers solar and wind power. Each subagent produced thorough, well-sourced coverage of its assigned topic. The web search subagent returned relevant results for every query it received. What is the most likely root cause?

A. The web search subagent used queries that were too narrow, missing results for other energy types
B. The coordinator’s decomposition named only solar and wind, so no agent was assigned the rest
C. The synthesis subagent failed to identify gaps in the research and request additional coverage
D. The document analysis subagent did not have access to sources covering other renewable energy types

Answer & explanation

Correct: B

B is correct because the coordinator is responsible for task decomposition. If it only assigns solar and wind as subtopics, no downstream agent can cover the missing categories. The root cause is always the coordinator’s decomposition when the output is incomplete in scope.
A is wrong because the web search subagent researched exactly what it was assigned and returned relevant results. The issue is what it was asked to search for, not how it searched.
C is wrong because the synthesis agent works with the research it receives. It cannot synthesise topics that were never researched. Gap identification during iterative refinement is the coordinator’s responsibility.
D is wrong because source availability is not the issue. The coordinator never asked any agent to research these other energy types. Even with perfect source access, unassigned topics would remain uncovered.

Question 2

Which of the following is a benefit of routing ALL subagent communication through the coordinator?

A. It reduces the total number of API calls required to complete a task
B. It allows subagents to share memory and build on each other’s findings automatically
C. It enables subagents to run in parallel without coordination overhead
D. It provides observability, consistent error handling, and controlled information flow

Answer & explanation

Correct: D

D is correct because centralised communication through the coordinator provides three specific benefits: observability (log and monitor every message in one place), consistent error handling (the coordinator applies uniform recovery policies), and controlled information flow (the coordinator decides what context each subagent receives).
A is wrong because routing through the coordinator may actually increase API calls compared to direct communication. The benefit is control and visibility, not efficiency.
B is wrong because subagents do NOT share memory. Communication through the coordinator means the coordinator explicitly passes information — there is no automatic memory sharing.
C is wrong because parallel execution is achieved through multiple Task tool calls in a single response, not through the routing pattern. Centralised routing is about control, not parallelism.

Question 3

A developer proposes allowing subagents to communicate directly with each other to reduce latency. Why is this approach problematic?

A. It breaks observability, consistent error handling, and controlled information flow
B. Direct communication is not supported by the Claude API
C. Subagents cannot process messages from other subagents
D. It would require every subagent to have access to all the other subagents’ tools, widening each one’s scope

Answer & explanation

Correct: A

A is correct because direct inter-subagent communication bypasses the coordinator, breaking the three benefits of hub-and-spoke architecture: observability (messages are no longer centrally logged), consistent error handling (no uniform recovery policies), and controlled information flow (no gatekeeper deciding what context flows where).
B is wrong because the issue is architectural, not a technical API limitation. The pattern is wrong regardless of whether the API supports it.
C is wrong because subagents can process any text input. The issue is not message processing capability but the loss of centralised control.
D is wrong because tool access is a separate concern from communication routing. Direct communication does not inherently require shared tool access.

Question 4

A coordinator invokes a web search subagent twice for two different subtopics. The second invocation returns results that contradict the first. What explains this?

A. The web search tool has a caching bug that returns stale results
B. The coordinator passed conflicting instructions to the two invocations
C. Subagents share no memory, so each invocation is fully independent
D. The second subagent inherited stale context from the first invocation

Answer & explanation

Correct: C

C is correct because subagent isolation means each invocation is completely independent. The second invocation has no knowledge of what the first found. Contradictory results are possible because each invocation searches independently and may find different sources.
A is wrong because the explanation lies in subagent isolation, not tool caching. Each invocation is independent by design.
B is wrong because while conflicting instructions could cause different results, the question asks what explains the contradiction — subagent isolation (no shared memory) is the architectural reason.
D is wrong because subagents do NOT inherit context from previous invocations. This answer contradicts the isolation principle.

Question 5

A multi-agent system’s output consistently misses entire categories of a broad topic. A developer proposes adding three more specialist subagents to increase coverage. Will this fix the problem?

A. Yes, more subagents means more comprehensive coverage of the topic
B. No, narrow decomposition simply yields equally narrow assignments
C. Yes, as long as the new subagents have access to different data sources
D. No — the fix is to allow subagents to communicate directly so they can identify gaps themselves

Answer & explanation

Correct: B

B is correct because the root cause is the coordinator’s task decomposition. If the coordinator only generates narrow subtopics, additional subagents receive equally narrow assignments. The fix is improving how the coordinator decomposes topics to ensure full breadth coverage.
A is wrong because more subagents do not help if they receive narrow assignments. Coverage depends on what is assigned, not how many agents are available.
C is wrong because data source access is irrelevant when the coordinator never assigns the missing categories. The agents cannot research topics they are never asked to investigate.
D is wrong because direct subagent communication breaks hub-and-spoke architecture and does not address the decomposition problem. Gap identification is the coordinator’s responsibility during iterative refinement.