CCA-F start here · Trap gym · Subagent context

Subagent context

A subagent ignores what the user said, repeats research that was already done, or loses the sources. The options offer shared memory, a stricter system prompt or a shorter summary. The trap is forgetting that a subagent starts with an empty memory: if a fact isn't in its Task prompt, it doesn't know it.

D1 Agentic ArchitectureExam guide 1.3About 12 minutes

At a glance: a subagent knows only what its prompt tells it, so pass findings and sources explicitly, and launch parallel subagents as several Task calls in one response.

Start here: the contractor who missed the meetings

You hire a contractor to finish a room. They were not at last week's meetings, so they don't know the client hates grey. If you don't write it in their brief, they will paint it grey. And if your brief says “use the paint we agreed”, they still don't know which paint.

A good brief says what to do, every decision already made, and where each fact came from: “Room 2, sage green (client email, 3 May), no grey.” The contractor can start at once and you can check their work.
Several rooms at once. If three rooms don't depend on each other, you hand out three briefs on the same morning. You don't wait for the first room to finish before writing the second brief.

Claude's multi-agent systems work the same way. The coordinator is the agent that talks to the user and splits the work. It hands a job to a subagent through the Task tool, and the text it writes in that call is the Task prompt. The subagent has isolated context: it does not see the coordinator's conversation, the user's rules or another subagent's results unless the Task prompt spells them out. And the coordinator can only use the Task tool if its allowedTools list (the tools the agent is permitted to call) includes "Task".

Coordinator its own history Chair: leave out trials under 200 patients Research findings: 3 trials, with sources allowedTools must include "Task" history: not inherited Task prompt the only channel Efficacy writerstarts empty Safety writerstarts empty Cost writerstarts empty several Task calls in one response
The one question to ask: was the fact the subagent got wrong actually in its Task prompt?

Words in the question that tell you which side you're on

Write it into the Task prompt

  • “ignores what the user said”
  • “repeats work already done”
  • “loses the sources”, “can't trace a figure”
  • “doesn't know the earlier findings”
  • “the issue we've been discussing”

Send parallel Task calls

  • “subagents run one after another”
  • “independent searches”
  • “takes too long”, “most of the time goes on waiting”
  • “no search needs another's results”

Try it: brief the writers

A drug-review committee asks for a summary of Dravelin, a made-up migraine drug. In earlier turns the chair told the coordinator to leave out trials under 200 patients, and a research subagent found three trials, each with its source. Now the coordinator hands the writing to three subagents: efficacy, safety and cost. Choose what goes in each Task prompt, how they are spawned, and whether the coordinator may use the Task tool at all.

1. What the Task prompt contains
2. How the three writers are spawned
3. The coordinator's allowed tools
coordinator writes a Task callsubagent, good resultsubagent, flawed result

The drug, trials, sources and times are made up and illustrative, chosen to show the shape. They are not measurements. The exam guide's points are that subagent context must be passed explicitly in the prompt, that complete findings go in a structured format separating content from metadata, and that parallel subagents are spawned as several Task calls in one response.

Things to try: Nothing: “use what we found above”, to see that a subagent has no “above”. Then Just the task: it searches again and brings back the excluded trial. Then findings as prose: the facts survive but the sources don't. Last, structured findings + one Task call per turn, then switch to parallel and watch the clock.

The traps

TrapWhat the wrong option looks likeWhy it loses
Doesn't existGive the subagents shared memory, or let them read the coordinator's historyWorks in some other frameworks, but here subagents inherit nothing automatically. The prompt is the only channel.
Wrong homePut this user's rules in the subagent's system promptFine for rules that never change. Facts that belong to one request, like this chair's size rule, go in that request's Task prompt.
Loses informationSummarise the earlier findings in one short sentence to save tokensThe guide asks for complete prior findings. A one-line summary drops the numbers and the sources.
Wrong shapeLet the subagents talk to each other directlyWorks in a mesh design, but the hub-and-spoke pattern keeps all communication going through the coordinator.
Waiting in a queueOne Task call per turn for jobs that don't depend on each otherSame results, several times the wait. Several Task calls in one response run them in parallel.

Practice: three exam-style questions

Answer first, then open “Think it through”.

Rule to remember

If it isn't in the Task prompt, the subagent doesn't know it. Send complete findings with their sources kept apart, and send independent jobs as several Task calls in one response.

Exam guide 1.3: “Subagent context must be provided explicitly in the prompt (no automatic inheritance / shared memory).” and “Include complete prior-agent findings in the subagent prompt; use structured formats to separate content from metadata (URLs, doc names, page numbers) for attribution; spawn parallel subagents by emitting multiple Task calls in one response”. Also 1.3: “Task tool spawns subagents; allowedTools must include "Task".”

← All traps