iterative-retrieval

Refining codebase context for subagents in large repositories involves a progressive 4-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. It scores relevance and refines search patterns, keywords, and exclusions across up to three cycles.

Handling context limits in multi-agent workflows is best achieved by iteratively retrieving and refining context. This approach uses a loop to evaluate relevance and refine queries, ensuring subagents receive only the most critical information without exceeding token limits.

Subagents fail during code exploration when initial search context lacks relevance. Resolve this missing-context failure by implementing an iterative retrieval loop that evaluates results and progressively refines search patterns, keywords, and exclusions across multiple cycles.

Initiate a DISPATCH with broad search patterns, EVALUATE the results, REFINE the criteria based on relevance scores, and LOOP up to three times. This progressively assembles highly relevant context for subagents.

Iterative retrieval works for context-limited tasks in large codebases by scaling search patterns across cycles. It evaluates and refines keywords and exclusions up to three times, ensuring subagents receive high-relevance context without exceeding limits.