Never use, launch, or delegate work to a sub-agent unless the user explicitly requests sub-agent usage for the current task.
Do not decide on your own that a task would benefit from sub-agents, regardless of its size, complexity, duration, or parallelizability.
If the user does not mention sub-agents, perform the task normally without them.
When the user explicitly requests sub-agents:
- If they specify which sub-agent, agent, harness, or model to use, follow that choice.
- If a required selection is missing, ask a short clarification before launching the sub-agent.
- Do not infer the choice from task type, language, repository, complexity, previous usage, or historical preferences.
- Previous sub-agent selections do not carry over unless the user explicitly says to reuse them.
- Do not launch additional sub-agents beyond the scope the user requested without explicit authorization.
This applies to implementation, review, research, debugging, testing, planning, investigation, and any other delegated work.
For work performed directly without sub-agents, reasonable defaults may be used for tools, skills, execution modes, reasoning levels, environments, or other implementation details unless the user explicitly specifies otherwise.
Do not interrupt the user with unnecessary selection questions when no sub-agent delegation is involved.
fast mode or equivalent accelerated model execution is OFF by default.
Enable it only when the user explicitly requests fast for that execution.
Do not infer permission from requests such as:
- "Do this quickly."
- "Use the fastest model."
- "Keep it short."
- "Use low reasoning."
- "Use the same model as before."
A previous fast choice does not automatically carry over.
No explicit fast request = fast OFF.
No explicit sub-agent request = no sub-agents.
Explicit sub-agent request + missing required selection = ask the user.
Direct execution may use reasonable defaults unless the user specifies otherwise.
No explicit fast request = fast OFF.