iterative-retrieval

To solve the subagent context problem, use iterative retrieval to progressively refine context through a 4-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. This balances precision and efficiency by selecting high-relevance files and identifying gaps.

The best way to manage token efficiency is using relevance scoring to evaluate candidate files during the EVALUATE phase. This identifies high-value sources and reduces context waste before spawning subagents with strict token limits.

The iterative retrieval loop refines context by dispatching an initial broad query, evaluating results for relevance, and then updating search patterns and keywords. This guided refinement closes information gaps within up to three cycles.

Use progressive context retrieval when spawning subagents for codebase exploration that start with incomplete or ambiguous project context. It is specifically designed for workflows facing context too large or missing context failures during task execution.

A limitation of the iterative retrieval pattern is its cycle-limited refinement, capping at up to three cycles. This constraint balances precision and efficiency but may leave information gaps if the codebase context is exceptionally complex.