iterative-retrieval

Multi-agent context overload is prevented by applying a four-stage loop—DISPATCH, EVALUATE, REFINE, LOOP—that prunes candidate files and assembles high-relevance context across a codebase within a maximum of three cycles.

Progressive retrieval is a pattern that evaluates file relevance against a task and refines queries iteratively, enabling context-driven collaboration between agents for bug fix analysis, feature implementation, and library exploration.

You refine search queries by running iterative retrieval cycles that dispatch searches, evaluate candidate files against the task, and refine criteria up to three loops to collect high-relevance files and discover repository patterns.

Yes, iterative retrieval works for subagents with limited initial context by progressively refining context to bridge knowledge gaps, locating relevant files, patterns, and terminology across large repositories and documentation.

The maximum number of cycles for iterative retrieval context refinement is three, ensuring the EVALUATE and REFINE stages converge efficiently without causing excessive context accumulation or misalignment.