Subagents: context isolation and orchestration for Claude

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. An agent must review 200 changelog files and produce one consolidated summary. A single loop fills its context with the raw file text long before it finishes reading, and you want each file read handled outside the main loop. Which approach should you take?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: D is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   Your answerA.Raise max_tokens on the main loop's requests to 128,000.max_tokens is an enforced ceiling on what the model generates in one response. The problem here is input: 200 files of source text consumed as context. Raising the output cap does not create room for them.Your answerB.Set output_config.effort to "max" on the main agent.output_config.effort controls reasoning depth and token spend, from low through max. Raising it makes the same overloaded loop think harder and cost more; it does not move the file text out of the context window.Your answerC.Disable adaptive thinking so the loop spends fewer tokens.Thinking blocks are not what fills the context here — the file contents are. Disabling thinking on Claude Opus 5 also carries two known failure modes, including tool calls emitted as visible text instead of tool_use blocks.The answerD.Use a multiagent session that delegates file reads to subagents.Work that fans out naturally across many files is the documented parallelization pattern for a multiagent session: fan out independent subtasks such as analyzing separate files and have the coordinator synthesize the results.

   Show the full explanation and doc quotes

   Because each delegated agent runs in its own context-isolated session thread with its own conversation history, the bulk file text stays in the subagents' contexts and only their conclusions return to the coordinator. The coordinator is left holding the synthesis while the delegated agents carry the reading.

   Written from platform.claude.com/docs/en/managed-agents/multiagent-orchestration

2. Your doc-reviewer subagent must apply your team's house-style skill. Your main session had that skill loaded when it delegated, but the subagent's review ignores every convention in it, and the subagent's transcript shows it never opened the skill at all. How should you restructure the delegation so the subagent applies that skill?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: A is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   The answerA.List the skill in the agent definition's skills field, which preloads it when the subagent startsSkills do not travel with a delegation. A subagent does not receive preloaded skill content unless the skill is named in AgentDefinition.skills, whose entries are preloaded at startup; unlisted skills remain invocable through the Skill tool but nothing loads them automatically.Your answerB.Load the skill in your main session immediately before delegating, so it is in context at the handoffWhat the parent has loaded is part of the parent's context, and a subagent receives none of it. Preloaded skill content is listed among the things a subagent does not inherit.Your answerC.Add the skill's directory to settingSources so the subagent loads it alongside the project CLAUDE.mdsettingSources is how project CLAUDE.md reaches a subagent. Skills are selected by the definition's skills field, which is a separate preload list.Your answerD.Set the agent definition's memory to 'project' so the skill's content is restored with project memoryConfuses memory, which selects a memory scope of 'user', 'project' or 'local', with the preload list that decides which skills start loaded

   Show the full explanation and doc quotes

   Naming the house-style skill in skills is what puts its content in the subagent's window from its first turn, which is what a convention the review must follow throughout requires.

   Written from code.claude.com/docs/en/agent-sdk/subagents

3. You want a repository investigation delegated to a built-in subagent so its file reads stay out of your main context. Afterwards you must put follow-up questions to that same agent about what it found, from a second query() call, without it starting the investigation over. Which approach should you take?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: B is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   Your answerA.Delegate to the built-in Explore agent, then resume it by passing its agent id and the captured session_id in the second call.Explore is one-shot. Its result carries no agentId, so there is no identifier to pass back, and a second call simply starts a fresh investigation.The answerB.Delegate to the built-in general-purpose agent, capture the agentId from its result, and resume that session in the second call.Of the three built-in subagents, only general-purpose is resumable. Explore and Plan are one-shot and do not return an agentId, so there is no handle to continue them with.Your answerC.Delegate to the built-in Plan agent, which inherits the main conversation's model, then question its findings in the second call.Plan is the read-only built-in used in plan mode. It inherits the main conversation's model, but like Explore it is one-shot and returns no agent id to resume against.Your answerD.Define a custom agent that preloads a repository-analysis skill, and invoke it again with the same description in the second call.Each subagent invocation creates a new instance rather than continuing an earlier one. Invoking the same definition again gives you a fresh conversation that has to redo the reading, whatever skills it preloads.

   Show the full explanation and doc quotes

   Resuming a subagent means capturing the parent's session_id, parsing agentId: <id> out of the Agent tool result, then passing resume: sessionId in the second query and naming the agent id in the prompt, because each query() call otherwise starts a new session.

   Written from code.claude.com/docs/en/agent-sdk/subagents