iterative-retrieval

Subagent context discovery is handled by progressively refining context retrieval through a four-phase loop of DISPATCH, EVALUATE, REFINE, and LOOP, allowing agents to identify relevant files, patterns, and terminology before task execution.

Iterative retrieval is a pattern for progressively refining context retrieval to solve the subagent context problem, enforcing a structured workflow that narrows down relevant files and terminology through controlled cycles.

You can use iterative retrieval by executing a retrieval cycle to gather and progressively refine relevant files for a given task, repeating the DISPATCH, EVALUATE, and REFINE phases up to three cycles to produce high-relevance results.

Yes, the iterative retrieval pattern controls context growth by limiting shared data per cycle to respect context constraints, ensuring subagents do not exceed their operational boundaries during multi-agent workflows.

You should use an iterative retrieval loop when subagents start with limited context and need to discover which files, patterns, and terminology are relevant, particularly during debugging, codebase onboarding, and feature research in large repositories.

The iterative retrieval pattern enforces a maximum of three cycles to produce high-relevance results, limiting the depth of context refinement if the codebase requires extensive discovery beyond the controlled phases.