When multi-agent designs pay off, how to brief a subagent, what it should return, and how to handle parallel work and conflicting findings.
In 30 seconds
Use subagents when work is broad, parallelisable and too big for one context window — not for small tasks.
A subagent starts fresh, so the brief must carry the objective, scope, sources, output format and stop rule.
Subagents return compact, structured findings with sources, never raw transcripts.
Independent subtasks run in parallel; dependent ones run in order.
Fan-out with briefs, fan-in with compact findings. Each subagent explores in its own context window.
Why split the work at all?
A single agent researching five sources fills its context window with search results, page text and dead ends. A coordinator can instead send each subagent to explore one area in its own context, then receive only the distilled result. The coordinator stays focused on the overall question.
The cost is real: more agents means more tokens and more coordination. For a short task with one tool, a single agent — or a fixed workflow — is simpler and cheaper. Reach for multiple agents when breadth and context size demand it.
Writing a brief that works
A fresh subagent knows nothing about the conversation that led to its task. "Research competitor pricing" produces vague results. A useful brief reads more like a ticket:
Objective: list current list prices for plans of competitors X, Y, Z
Scope: public pricing pages only; ignore reseller sites
Output: JSON array of {company, plan, price, currency, source_url, quote}
Stop when: all three companies covered, or after 8 searches
What should come back
Each claim should travel with its source and the location of the supporting passage. That makes the final report verifiable and lets the coordinator spot disagreements. Raw transcripts defeat the purpose: they move the context problem from the subagent to the coordinator.
Parallel or sequential?
If sub-questions are independent, send the subagents out at the same time: total latency approaches the slowest one rather than the sum. Run steps in order only when a later step genuinely needs an earlier result.
When findings disagree
Keep both claims with their sources. Then either check the primary source with a focused follow-up or report the conflict openly. Averaging two numbers produces a figure that nobody reported, and quietly picking one hides the uncertainty from your reader.
Part of Domain 1: Agentic Architecture & Orchestration (27% of the exam). Explanations are our own, checked against Anthropic's documentation on November 22, 2026.
Common traps
Trap
Why it fails
Do this instead
Returning full search transcripts
The coordinator's context overflows
Return structured findings with sources
"Be thorough" as the brief
A fresh subagent lacks the context to know what thorough means
State objective, scope, sources, format and stop rule
Averaging conflicting figures
Creates a number no source reported
Report both with sources, verify or flag
Check yourself
Two original practice questions on this lesson. Try them before opening the answers.
Check yourself 1
Scenario: Multi-agent research system
A research system uses one coordinator and several subagents that each search a different source. Subagents return their full raw search transcripts, and the coordinator's context fills up after three subagents. Which change best addresses this?
A. Have each subagent return a compact, structured summary of findings with their sources
B. Increase the coordinator's max_tokens so it has more room to hold the incoming transcripts
C. Run every search inside the coordinator's own conversation so all evidence stays together
D. Cap each subagent at a single search query so its transcript is always short enough
Show answer and explanation
Answer: A. Context isolation is a key benefit of subagents: each explores in its own window and hands back only what the coordinator needs. Distilled, structured findings with source references keep the coordinator small while preserving traceability. Merging everything into one conversation defeats the purpose; max_tokens controls output length, not how much context is available; one query per subagent cripples research quality.
Check yourself 2
Scenario: Multi-agent research system
A coordinator delegates "research competitor pricing" to a subagent, and results come back vague and incomplete. Which change to the delegation is most likely to help?
A. Give the subagent access to every tool in the system so that it is never blocked from a source
B. Tell the subagent to be thorough and to keep searching until it is confident in its results
C. Give the subagent a brief with objective, scope, sources, output format and a stop condition
D. Have the coordinator run the research itself, since it already has the full context
Show answer and explanation
Answer: C. A fresh subagent starts without the coordinator's context, so vague delegation produces vague work. A good brief states the objective, boundaries, sources, expected output structure and when to stop. "Be thorough" is not actionable, broad tool access adds confusion and risk, and abandoning delegation gives up parallelism and context isolation.