iterative-retrieval

Iterative retrieval progressively narrows large codebase contexts by running a four-phase loop—DISPATCH, EVALUATE, LOOP, REFINE—across up to three cycles. It scores file relevance and dynamically refines queries to return only highly relevant files to subagents.

Evolving context in multi-agent workflows is handled by iteratively refining retrieval queries as subagents work. The process dispatches broad queries, evaluates results, and loops back with refined queries, ensuring subagents receive context they did not know they needed initially.

Iterative context retrieval manages token usage by filtering out irrelevant files through relevance scoring and dynamic query refinement. By executing a maximum of three retrieval cycles, it excludes unnecessary context, keeping the token count within size limits.

You need iterative retrieval when spawning subagents that require evolving codebase context. It is specifically designed for scenarios where subagents do not know what context they need until they start working, or when large contexts must be narrowed without losing critical information.

Yes, iterative retrieval filters codebase files by applying relevance scoring during the EVALUATE phase. It dynamically refines search queries and excludes irrelevant files over multiple cycles, returning a highly relevant subset of files for your subagent tasks.

The retrieval loop is limited to a maximum of three cycles, meaning context refinement stops after the third iteration regardless of whether optimal relevance is achieved. This bounds execution time but may miss deeply nested context in extremely large codebases.