sadd:launch-sub-agent

To launch a sub-agent with isolated context, you provide a discrete task description. The system automatically selects an appropriate model, applies specialized agent prompts, and enforces a self-critique verification loop before dispatching the final result.

Model selection for sub-agent dispatch works by analyzing task complexity and expected output size, automatically choosing between opus, sonnet, and haiku to ensure efficient resource allocation and optimal performance for research or testing tasks.

Zero-shot chain-of-thought is a structured reasoning prefix prepended to sub-agent tasks. It improves agent dispatch by enforcing stepwise reasoning before execution, which reduces context pollution and enhances the reliability of the final output.

Yes, you can use automated sub-agents for documentation edits and architecture design. The system matches domain-specific agent prompts to the task, spawning a haiku agent for trivial edits or an opus agent for complex architecture.

A sub-agent requires a self-critique verification loop to ensure accuracy and reliability. This mandatory loop asks targeted questions and checks evidence before accepting results, preventing unverified outputs from polluting the main context.

Automated model selection for sub-agent orchestration relies on complexity heuristics, meaning highly ambiguous or deeply nested cross-domain tasks might require manual model overrides to ensure the correct agent context and specialized prompts are applied.