What you need to know
In a coordinator-subagent (hub-and-spoke) system, one coordinator agent sits at the centre and specialised subagents sit at the spokes. The coordinator owns four jobs:
- Decompose the request into pieces of work.
- Delegate each piece to the right subagent.
- Aggregate what comes back into one result.
- Decide which subagents a given request needs at all.
Subagents do not talk to each other. Every message goes through the coordinator, which gives you one place for logging, error handling and control over what information flows where.
Subagents start with a clean context
A subagent does not inherit the coordinator's conversation history. It sees its own system prompt and whatever the coordinator puts in the task it hands over, nothing more. Lesson 1.3 covers how to pass context; here the point is architectural: if the synthesis subagent needs the search findings, the coordinator must send them.
Match the pipeline to the query
A coordinator that always runs every subagent wastes time and money. "When did the regulation take effect?" needs one search, not search plus document analysis plus synthesis plus a formatted report. A good coordinator analyses the query first and invokes only the subagents it needs: one for a simple fact, the full pipeline for a broad research question.
Decomposition can fail by being too narrow
The commonest quality failure in research systems is not a bad subagent but a narrow decomposition. Asked about "AI in creative industries", a coordinator that splits the work into digital art, graphic design and photography never asks anyone about music, film or writing. Every subagent does its job well and the report is still incomplete. The fix is at the coordinator: a broader decomposition, and a review of coverage before the report is written.
Partition to avoid duplication
When several subagents search in parallel, give each a distinct slice: separate subtopics, separate source types, separate time ranges. Overlapping assignments burn tokens finding the same sources twice. Where topics genuinely intersect, assign the overlap to one subagent deliberately and deduplicate at synthesis.
Iterate until coverage is sufficient
A single pass rarely produces complete coverage. The stronger design is an iterative refinement loop:
- Subagents research; synthesis produces a draft.
- The coordinator evaluates the draft against the question: which subtopics are thin or missing?
- It re-delegates targeted queries for those gaps to the search and analysis subagents.
- Synthesis runs again with the new findings. Repeat until coverage meets the bar.
In the exam
Scenario 3 (multi-agent research) is the natural home for these questions. Look for: subagents communicating directly, a coordinator that always runs everything, a report with whole missing subtopics, duplicated work across subagents, or a single-pass design with no gap check. The right answer usually changes the coordinator's behaviour.
Check your understanding
A broad report covers three subtopics in depth and never mentions four others. Logs show each subagent did its assigned work well. Where is the defect?
Show answer
Answer: C. If the missing subtopics were never assigned, no subagent could cover them. Narrow decomposition is a coordinator failure, and the fix belongs there.
Two subagents have started sending each other intermediate results directly. What is the main architectural cost?
Show answer
Answer: A. Hub-and-spoke works because all communication passes through the coordinator. Side channels create traffic nobody monitors and failures nobody handles consistently.
Simple factual questions cost almost as much as full reports. What should change?
Show answer
Answer: D. The coordinator should assess each query and invoke only what it needs. A simple fact needs one search, not the whole pipeline.
Apply it: Redesign Whitfield's coordinator
Whitfield's coordinator runs all four subagents for every question and produces reports with missing subtopics. Write the coordinator's operating rules.
- Write a routing rule. Say how the coordinator decides between a single search and the full pipeline, with one example of each.
- Write the decomposition step. For the question "How will EU battery regulation affect the e-bike market?", list the slices you would assign and how you keep them from overlapping.
- Add the gap loop. Describe what the coordinator checks in a draft, what it does when a subtopic is thin, and when it stops.
Done when: the loop has a stopping condition that is about coverage, not a fixed number of rounds. - State the communication rule between subagents.
Show model answer
1. Routing. The coordinator first classifies the question. A lookup (one fact, one date, one number) gets a single web search task and a short answer with its source. A research question (trends, comparisons, forecasts) gets the full pipeline. Example lookup: "When did the EU battery regulation enter into force?" Example research: "How will it affect e-bike makers' costs to 2030?"
2. Decomposition
| Slice | Subagent | Sources |
|---|---|---|
| What the regulation requires and when | Web search | Official texts and regulator summaries |
| Supply chain and recycling capacity | Web search | Industry bodies and news, last 24 months |
| Pricing and margins | Document analysis | Client-supplied cost data |
| Demand and market size | Web search | Market studies, with publication dates |
Each slice is defined by subtopic and source type, so two subagents should not be reading the same pages. Recycling costs touch both the supply chain and pricing slices; it is assigned to supply chain only, and synthesis merges duplicates.
3. Gap loop. After each draft the coordinator lists the question's subtopics and marks each as covered (two or more independent sources), thin (one source) or missing. For each thin or missing subtopic it sends one targeted query to the right subagent, then re-runs synthesis. It stops when every subtopic is covered or when a subtopic has been retried twice without new sources, in which case the gap goes into the report's coverage section.
4. Communication. Subagents never message each other. Everything they produce returns to the coordinator, which decides what each next task receives.