0 Subagent Workflow

Subagent Workflow

Use subagents to reduce hallucination and keep main context clean.

Spawn read-only subagents when the task is broad, uncertain, or cross-file:

1. Searcher: find relevant files, symbols, call sites, and existing patterns.
2. Skeptic: identify edge cases, risky assumptions, bugs, and behavior regressions.
3. Verifier: find tests, fixtures, commands, and the smallest validation path.

Cost control:

• Searcher: force low reasoning; do not inherit the main thread model/reasoning.
• Verifier: force low reasoning; do not inherit the main thread model/reasoning.
• Skeptic: force medium reasoning; do not inherit xhigh from the main thread.
• If the subagent tool cannot choose a cheaper reasoning level, do not spawn a read-only subagent for simple search; use local tools in the main thread.
• Use high/xhigh subagents only when explicitly asked or needed for deep architecture reasoning.

Each subagent returns only:

• conclusion
• file:line evidence
• confidence level
• unknowns
• recommended next action

Rules:

• Subagents do not edit files.
• Do not return raw logs unless essential.
• Wait for all summaries before implementing.
• Main agent cross-checks important claims against source files before editing.
• Main agent owns final edits, verification, and closeout.