Skip to content

Task Type C — Large Exploration

When to Use

Use when the task is open-ended, you don't know the structure yet, or a full Type B read would exceed ~30% of remaining context. Use Type B when the unit is bounded enough for main context.

Decision

Signal Action
Open-ended "how does X work" across many files Dispatch Explore subagent
You don't know the structure or entry points yet Dispatch Explore subagent
Full read would exceed ~30% of your remaining context Dispatch subagent (Explore, general-purpose, or specialized)
Research question spans multiple repos or unfamiliar domains Dispatch subagent

Pattern

1. Frame the research question for the subagent
   - What are you trying to understand?
   - What structured output do you want back?
     (e.g., "list of classes with their methods and relationships")

2. Dispatch to Explore or general-purpose subagent
   - The subagent burns ITS context doing full reads
   - It returns a structured summary to yours

3. Use the summary to decide what to Read targeted in your main context
   - Usually only 2-3 files from the original N

The subagent trades its context for yours. It can do 20+ full reads and return a 2-page summary, preserving your main conversation for actual authoring or implementation.

Common Mistakes

  • Wrong: Reading 20 files in main context to "get a feel" for the codebase → Right: Delegate to Explore; it does the full reads and returns a structured index
  • Wrong: Asking the subagent "summarize" without specifying what fields you need → Right: Ask for specific structured output (class names, method signatures, inheritance chains, config keys)
  • Wrong: Re-dispatching to the subagent for every follow-up → Right: Do targeted Reads in main context for specific files the summary flagged as important

See Also