Sub-agents
Sub-agents handle independent analysis or edits and return results to the main task. They differ from conversation branches, which copy history.
Open
Configure and enable
Enable orchestration and configure usable provider/model candidates in Settings. The switch alone is insufficient without enabled models. Start with a read-only code review.
Divide work
Read-only tasks
Split UI logic, data flow or test coverage into concrete questions with reading scope and expected output. Avoid duplicated investigation.
Tasks that write
Declare non-overlapping write paths. Read-only tasks cannot also declare writes, and writable tasks require a write scope. Coordinate shared files and interfaces through the main task.
Example
Have sub-agents independently inspect UI event handling and persistence without edits. Return file-specific causes. The main task should consolidate findings, implement the fix and run relevant checks.
Compare reports with actual code and inspect the final diff. A successful sub-agent report does not prove files are saved or tests pass.
Limits and diagnosis
Sub-agents cannot recursively delegate. Check orchestration, candidates, tool switches and path scope when unavailable. Reread files after write conflicts. Extra calls affect usage; see tool integration.