Core Concepts
Subagents
Subagents are delegated helper runs that reduce main-session context load and can execute in parallel.
What subagents are
Subagents are spawned by the main agent through the spawn-subagent tool for focused exploration, planning, review, or research tasks.
Use them when you need parallel investigation without polluting the main working context.
How they run
- Subagents can be spawned in parallel for independent tasks.
- They use a restricted, read-first toolset and are intended for analysis, not direct edits.
- They cannot spawn nested subagents.
- Explore subagents use a fast model profile; most other subagent types use a heavier profile for depth.
Best use case
Delegate discovery work (find files, compare options, gather evidence), then implement in the main session with a cleaner context window.
Limits and tradeoffs
- Subagent output quality depends on task framing and provided context.
- Very vague delegation requests can produce shallow findings.
- Use explicit expected output format when you plan to merge multiple subagent results.