← Study Guide
3Exam Scenario

Multi-Agent Research System

A coordinator delegates to specialized subagents (web search, document analysis, synthesis, report generation) to produce comprehensive, cited reports.

Agentic Architecture & OrchestrationTool Design & MCP IntegrationContext Management & Reliability

What this scenario tests

Coverage gaps trace to decomposition

If reports cover only visual arts and miss music/writing/film while every subagent 'succeeded', the coordinator's decomposition was too narrow. Fix the coordinator, not the correctly-working downstream agents.

Error propagation on timeouts

A search subagent timeout should return structured error context (failure type, attempted query, partial results, alternatives). Don't return a generic 'unavailable', don't mark failure as empty success, and don't kill the whole workflow.

Scoped cross-role tools

85% of synthesis verifications are simple fact-checks. Give synthesis a scoped verify_fact tool for those; keep complex verifications routed through the coordinator. Don't hand it all web tools (over-provisioning) or rely on speculative caching.

Provenance through synthesis

Each finding carries claim, evidence excerpt, source URL/name, and publication date. Synthesis must preserve attribution; conflicting stats are annotated with sources and dates, not silently reconciled.

Parallel delegation

Emit multiple Task calls in one coordinator response for parallel subagents; allowedTools must include Task; subagents receive findings explicitly (no inherited context).

Edge Cases & Traps

  • Partition subtopics/source types across subagents to minimize duplication.
  • Run iterative refinement: evaluate synthesis for gaps, re-delegate targeted queries, re-synthesize.
  • Distinguish access failure (retry) from valid empty results (no matches).
  • Annotate coverage gaps in the final report when sources were unavailable.

🧠 Knowledge Check · 15 questions

Score: 0/15

1.Reports miss whole subtopics though all subagents succeeded. Where do you fix it?

2.85% of the synthesis agent's verifications are simple fact-checks; 15% need deep investigation. Best way to cut round-trip latency?

3.How should findings be structured so citations survive synthesis?

4.Which is required for a coordinator to spawn subagents in parallel?

5.A web-search subagent times out. What should it return to the coordinator?

6.Why is marking a failed search as an empty successful result dangerous?

7.Two subagents keep fetching the same sources. How do you cut duplication?

8.The first synthesis pass leaves gaps in coverage. Best coordinator behavior?

9.Two sources report conflicting statistics. How should synthesis handle it?

10.No delegation is happening — the coordinator never spawns subagents. First thing to check?

11.How should coordinator prompts be written to keep subagents adaptable?

12.Some sources were unavailable during the run. What should the final report do?

13.Someone proposes giving the synthesis agent all web-search tools 'just in case'. Why is that wrong?

14.A finding needs to keep its citation through synthesis. Which format preserves attribution?

15.How does the synthesis agent get access to the earlier web-search and document-analysis findings?