The coordinator's split
A multi-agent report comes back incomplete, wrong or slow, and the options offer to fix the subagents. Often the subagents did exactly what they were told. The trap is treating the symptom: blaming the subagent for a fault the coordinator made when it split and briefed the work.
D1 Agentic ArchitectureExam guide 1.2, 1.3 and 1.6About 15 minutes
Start here: the head chef and the order tickets
A head chef is catering a dinner. She talks to the client, then writes a ticket for each cook.
If the dinner has no main course, every cook still did their job: the chef never wrote a ticket for it. If the soup has bacon in it, the cook didn't know about the vegetarian; it wasn't on the ticket. Retraining the cooks fixes neither. In Claude's multi-agent systems the chef is the coordinator, the cooks are subagents, and the ticket is the Task prompt. A subagent starts with an empty context and knows only what that prompt says.
- Whole topics missing, every subagent fine: the coordinator's split was too narrow. Widen the decomposition.
- A subagent ignores something the user said: it never received it. Write it into the Task prompt; subagents inherit nothing.
- Independent searches run one after another: send several Task calls in one response so they run in parallel.
- Subagents follow a recipe and miss what matters: brief them with goals and quality criteria, not step lists.
Words in the question that tell you which side you're on
The coordinator's fault
- “all subagents complete successfully”, “within its topic”
- “the coordinator split the work into…”
- “the user told the coordinator…”, “agreed earlier”
- “no search needs another's results”
- “follow the steps exactly”
The subagent's own fault
- the subagent had the facts in its prompt and still got it wrong
- its results are poor for the topic it was given
- it uses tools outside its role
- only then: improve its prompt, examples or tools
Try it: replay the run
A coordinator researches “How is AI changing the creative industries?” for a museum board. Pick what went wrong, then pick a fix. The simulator replays the run: who works when, what each subagent was told, and what the report covers. The diagram above shows where your fix acts.
Times and counts are illustrative, chosen to show the shape. They are not measurements. The exam guide's points are that narrow decomposition leaves topics uncovered, subagents get context only through their prompt, parallel work goes out as several Task calls in one response, and briefs should be goals plus quality criteria.
Things to try: Whole topics missing + Improve the subagents, then + Widen the split. Then + Add a gap check: it works, but look at the clock. Last, The user's rules ignored + Improve the subagents: fewer old sources, never zero.
The three traps
| Trap | What the wrong option looks like | Why it's wrong |
|---|---|---|
| Blaming the subagent | Better search terms, more sources per subagent, a stricter subagent system prompt, a new tool | The subagents did what they were asked. The missing topic or fact was never in their Task prompt. |
| Catching it late | A gap check after synthesis, a reviewer, a filter that drops bad results | A good safety net, but it pays for a wasted round and leaves the cause in place for the next run. |
| Right layer, wrong lever | Parallel calls for a coverage gap; a wider split for missing context; more steps for a recipe that is too rigid | Split, brief and timing fix different symptoms. Match the lever to what went wrong. |
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 1.2: “Risk: overly narrow decomposition → incomplete coverage of broad topics.” Exam guide 1.3: “Subagent context must be provided explicitly in the prompt (no automatic inheritance / shared memory).” and “spawn parallel subagents by emitting multiple Task calls in one response; write coordinator prompts as goals + quality criteria, not step-by-step procedures.”