iterative-retrieval

Progressive retrieval patterns solve the missing context problem by incrementally gathering, evaluating, and refining relevant codebase context for subagents. This ensures subagents start with limited knowledge but progressively discover the appropriate materials needed to act effectively.

The best way to refine codebase search context for subagents is using a 4-phase iterative retrieval loop: DISPATCH, EVALUATE, REFINE, and LOOP. This pattern iteratively evaluates candidate files for relevance, identifies context gaps, and refines search criteria over up to three cycles to build high-relevance context.

Iterative retrieval works by dispatching initial search criteria, evaluating returned candidate files for relevance, and refining criteria to fill identified gaps. It operates in a loop with up to three cycles, progressively discovering code and terminology to build high-relevance context for subagents.

You need progressive context retrieval when spawning subagents that require evolving project context or when addressing context-too-large failures. It is ideal for multi-agent workflows where subagents must operate with progressively discovered code and terminology instead of static, potentially oversized context.

The iterative retrieval loop is limited to a maximum of three cycles when evaluating candidate files and refining search criteria. This bound prevents infinite searches but may limit exhaustive context discovery in exceptionally large codebases where deeper progressive refinement is required.