Multi-agent research mind maps
Everything the CCA-F exam asks about a coordinator and its subagents, on one page. It starts with which exam scenarios bring these questions. Then six mind maps, each in the form "what you see, where the fault is, the fix", with the rule to remember, the exam-guide line it comes from, and one practice question.
Domain 1 · 27%Domain 2 · 18%Domain 5 · 15%Guide 1.2, 1.3, 1.6, 2.2, 2.3, 5.3, 5.6About 30 minutes
Which scenarios ask about multi-agent systems?
The exam shows you 4 scenarios out of a bank of 6. One is built around a coordinator and subagents. Two others reuse parts of it:
| # | Scenario | Domains | Multi-agent questions? | Maps to study |
|---|---|---|---|---|
| 3 | Multi-Agent Research System | D1, D2, D5 | Yes, the core one. Coordinator splits the work across search, analysis, synthesis and report subagents. Reports need citations | All six |
| 1 | Customer Support Resolution Agent | D1, D2, D5 | Partly. Same Domain 1 and 2 ideas (tools per agent, structured errors), but one agent with hooks | 4, 5 |
| 4 | Developer Productivity with Claude | D2, D3, D1 | A little. The Explore subagent keeps noisy searching out of the main context | 2 |
| 2 | Code Generation with Claude Code | D3, D5 | No | — |
| 5 | Claude Code for Continuous Integration | D3, D4 | No | — |
| 6 | Structured Data Extraction | D4, D5 | No | — |
Guide §5, scenario 3: “Multi-Agent Research System — coordinator delegates to search/analysis/synthesis/report subagents; cited reports. (D1, D2, D5)”
First, the picture: hub and spoke
Most multi-agent questions come down to one thing: find which box the fault is in. If the question says the subagents “worked correctly”, the fault is usually in the coordinator.
The six maps at a glance
| # | Question | Usual answer | Practise |
|---|---|---|---|
| 1 | The report misses whole topics or repeats work | Fix the coordinator's split of the work | Coordinator split |
| 2 | A subagent doesn't know what it needs | Pass the context explicitly in its prompt. It inherits nothing | Subagent context |
| 3 | Fixed steps or adaptive? | Known steps: a fixed chain. Open-ended research: plan from what you find | Chain or decide |
| 4 | Which tools does each agent get? | Only the few its role needs, plus a scoped shared tool for frequent needs | Too many tools |
| 5 | A subagent failed. What goes back? | A structured error with the partial results, never silence or a full stop | Tool failure |
| 6 | Sources disagree or citations are lost | Keep a claim-to-source map. Show both figures with their sources and dates | Sources |
A yellow row in each map is the runner-up: the case where a different answer wins because one fact in the question changes.
Mind map 1: The report misses whole topics or repeats work
| What you see | Where the fault is | Fix |
|---|---|---|
| Every subagent did its part well, but whole areas are missing from the report | The coordinator split the topic too narrowly | Fix the coordinator's decomposition so it covers the whole question |
| Two subagents search the same ground | Their scopes overlap | Partition the scope so each has a clear, separate area |
| A simple question runs through every subagent | The full pipeline always runs | The coordinator picks only the subagents the query needs |
| The first draft has gaps | One pass and done | Refinement loop: check the synthesis for gaps, send new tasks, synthesize again |
| One subagent was given the right area but covered it badly | That subagent (its prompt or its tools), not the coordinator | Fix that subagent |
Guide 1.2: “Risk: overly narrow decomposition → incomplete coverage of broad topics. Dynamically select subagents rather than always routing the full pipeline; partition scope to minimize duplication; iterative refinement loops (evaluate synthesis for gaps → re-delegate → re-synthesize).”
- The search subagent's queries are too narrow
- The coordinator split the topic too narrowly
- The synthesis subagent dropped some findings
- The report subagent's template has too few sections
Mind map 2: A subagent doesn't know what it needs
| What you see | Where the fault is | Fix |
|---|---|---|
| The synthesis subagent ignores findings the coordinator already has | Subagents don't inherit the coordinator's history | Put the complete earlier findings in its prompt |
| Citations get lost between agents | Content and source details are mixed in free text | A structured format that keeps content apart from metadata (URL, document name, page) |
| The coordinator can't start any subagent | "Task" is missing from allowedTools | Add "Task" to the coordinator's allowed tools |
| Independent subagents run one after another and it's slow | One Task call per turn | Emit several Task calls in one response so they run in parallel |
| Step B needs step A's output | Running them in parallel gives B nothing to work with | Run them one after the other |
| Subagents follow the coordinator's step list too rigidly | The coordinator wrote a procedure | Write goals and quality criteria, not steps |
Guide 1.3: “Task tool spawns subagents; allowedTools must include "Task". Subagent context must be provided explicitly in the prompt (no automatic inheritance / shared memory). Spawn parallel subagents by emitting multiple Task calls in one response; write coordinator prompts as goals + quality criteria, not step-by-step procedures.”
- Tell the synthesis subagent to “always cite sources”
- Pass the findings to synthesis as structured records, each with its claim, URL and page
- Give the synthesis subagent the web search tool so it can find sources again
- Merge search and synthesis into one agent
Mind map 3: Fixed steps or adaptive?
| What you see | Where the fault is | Fix |
|---|---|---|
| The steps are known and the same every time (review each file, then check across files) | — | A fixed chain of prompts (prompt chaining) |
| Open-ended research where each finding changes what to look at next | A fixed plan can't follow new leads | Adaptive decomposition: create subtasks from what is found |
| One big pass over many documents gives shallow, contradictory results | Attention is spread too thin | One pass per document, then a separate combining pass |
| An adaptive agent is used for a routine job with the same steps every time | It adds cost and unpredictability for no gain | Use the fixed chain |
Guide 1.6: “Fixed sequential pipelines (prompt chaining) vs dynamic adaptive decomposition based on intermediate findings. Prompt chaining: analyze each file individually, then a cross-file integration pass. Adaptive investigation plans generate subtasks from what's discovered.”
- A fixed chain: search, analyse, synthesize, report
- A coordinator that creates new subtasks as findings come in
- One agent with all the tools and a long prompt
- Run the fixed chain twice and merge the results
Mind map 4: Which tools does each agent get?
| What you see | Where the fault is | Fix |
|---|---|---|
| One agent has 18 tools and picks badly | Too many tools | Split by role so each agent has 4 or 5 tools |
| The synthesis agent starts doing web searches itself | It has tools outside its role | Scope tools per agent (set them in each AgentDefinition) |
| Synthesis needs many quick fact checks, and each round trip through the coordinator slows things down | Every small check goes the long way | Give synthesis a scoped verify_fact tool for simple checks |
| The check is complex and needs real research | A small tool can't do it | Send it through the coordinator to the search subagent |
| Someone suggests giving synthesis all the search tools “to be safe” | Too much access | Keep least privilege. A scoped tool covers the common case |
Guide 2.3: “Too many tools (18 vs 4-5) degrades selection reliability. Agents with off-specialization tools misuse them. Scoped access; limited cross-role tools for high-frequency needs.” Guide 1.3: “AgentDefinition: descriptions, system prompts, tool restrictions per subagent.”
- Give synthesis all the search subagent's tools
- Give synthesis a scoped
verify_facttool for simple lookups, and keep sending complex checks through the coordinator - Collect every check and run them all at the end
- Cache extra context ahead of time
Mind map 5: A subagent failed. What goes back to the coordinator?
| What you see | Where the fault is | Fix |
|---|---|---|
| The coordinator gets only “search unavailable” and can't decide what to do | A generic status hides the details | Return a structured error: type of failure, the query tried, partial results, alternatives |
| A timeout comes back as an empty, successful result | The error is hidden | Report it as a failure. An access failure is different from a real empty result |
| One failed search stops the whole research run | It ends everything on a single failure | Let the coordinator recover: retry, try another source, or carry on with a note |
| A subagent sends every small error up to the coordinator | It escalates too early | Recover locally. Send up only what it couldn't fix, with the partial results |
| The search worked and there really are no results | Nothing failed | Return a successful empty result. It's not an error |
Guide 5.3: “Structured error context (failure type, attempted query, partial results, alternatives) enables coordinator recovery. Distinguish access failures (retry) from valid empty results. Generic statuses hide context; silently suppressing errors or terminating on single failure are both anti-patterns.” Guide 2.2: “local recovery in subagents, propagate only unresolved errors with partial results.”
- The 3 sources, marked as a full success
- An exception that stops the workflow
- An error saying it timed out, the query it ran, the 3 sources found, and other sources worth trying
- Retry quietly, then return “search unavailable” if it still fails
Mind map 6: Sources disagree or citations are lost
| What you see | Where the fault is | Fix |
|---|---|---|
| The final report has claims but no sources | Summarizing dropped the link between claim and source | Keep a claim-to-source map through every step |
| Two sources give different figures | The agent picked one and dropped the other | Report both figures, each with its source |
| The figures differ because one is from 2022 and one from 2025 | It's a change over time, not a contradiction | Require publication dates so the difference is explained |
| Financial figures are buried in paragraphs | Every kind of content is written as prose | Show financial data as tables and news as prose |
Guide 5.6: “Preserve structured claim-source mappings through synthesis; annotate conflicting statistics with source attribution (don't arbitrarily pick one). Require publication/collection dates so temporal differences aren't misread as contradictions; render content types appropriately (financial as tables, news as prose).”
- Use the government figure, since it's more official
- Average them to 15.5%
- Show both figures, each with its source and date, and note that they differ
- Leave the figure out because the sources conflict
Three traps that catch people
1 · “The subagents worked correctly”
When the question says this, it's telling you the fault is not in the subagents. Look at the coordinator: how it split the work, what context it passed, and how it combined the results.
2 · Subagents share nothing automatically
There's no shared memory between coordinator and subagent. An option that assumes a subagent “already knows” earlier findings is wrong. The fix is always to put them in its prompt.
3 · Task tool: the exam's answer vs today
The Agent SDK has been renaming things, but the guide says subagents are spawned with the Task tool, and "Task" must be in allowedTools. On the exam, follow the guide. More on Exam vs 2026, item 3.
How to study this
Related: Domain 1 hub · Domain 5 hub · Extraction mind maps · Claude Code mind maps · MCP mind maps · all traps · CCA-F start here