Orchestration Patterns

The CCAR-F exam focuses on exactly one multi-agent orchestration pattern: hub-and-spoke. A central coordinator (the hub) manages every interaction with specialized subagents (the spokes). The coordinator decomposes the task, selects which subagents to invoke, passes each one explicit context, aggregates their results, and applies uniform error handling. Subagents only ever talk to the coordinator - never to each other.

The cardinal rule: all communication flows through the coordinator

Subagents never communicate directly with one another - not for efficiency, not for convenience, not for any reason. Centralizing every message through the coordinator buys three things that a mesh of chatty agents cannot provide:

• Observability - one place to log and monitor every interaction.
• Consistent error handling - the coordinator applies uniform recovery policies.
• Controlled information flow - the coordinator decides exactly what context each subagent receives.

Subagent isolation

This is the single most misunderstood idea in multi-agent systems. Subagents do not automatically inherit anything from the coordinator. There is no shared memory, no global state, and no cross-invocation history.

• No automatic context inheritance - subagents do not receive the coordinator's system prompt, its conversation history, or other subagents' results.
• No cross-invocation memory - calling the same subagent twice gives the second call zero knowledge of the first.
• Explicit context passing is mandatory - every fact a subagent needs must be written into its prompt by the coordinator.

What the coordinator owns

The narrow-decomposition failure mode

This is the pattern the exam tests most. Suppose a system researching "the impact of AI on creative industries" produces a report thorough on visual arts but silent on music, writing, and film. The instinct is to blame the search or synthesis subagents - but they performed perfectly on what they were assigned. The failure originates in the coordinator's decomposition. The coordinator only assigned visual arts, so nothing else was ever researched.

# Coordinator decomposes BREADTH-FIRST, then partitions scope.
subtopics = decompose(topic)  # e.g. ["solar", "wind", "geothermal",
                              #       "tidal", "biomass", "fusion"]

# Each subagent gets a DISTINCT, non-overlapping assignment plus
# explicit context - it inherits nothing automatically.
findings = []
for sub in subtopics:
    result = coordinator.invoke_subagent(
        agent="web_search",
        prompt=f"Research {sub} within: {topic}. Context: {shared_context}",
    )
    findings.append(result)

# Iterative refinement: re-delegate for any gaps before synthesizing.
while has_coverage_gaps(findings, subtopics):
    findings += coordinator.refill_gaps(findings, subtopics)