iterative-retrieval

Iterative retrieval automates context gathering for multi-agent workflows by progressively refining retrieval results. It uses a 4-phase loop—DISPATCH, EVALUATE, REFINE, LOOP—to assemble high-relevance context files for agent activation.

Reducing token usage during codebase exploration is achieved through iterative retrieval cycles that score relevance and identify gaps. This enforces staged discovery with a cap of three cycles to minimize wasted tokens.

Iterative retrieval works by dispatching a broad intent, evaluating candidate files for relevance, and refining the query up to three cycles. This loop progressively narrows down high-relevance context files needed for agent activation.

Yes, iterative retrieval is suitable for code exploration, debugging, feature work, and complex task orchestration across agents. It solves the subagent context problem by enabling staged discovery of necessary files.

The main limitation of this context optimization loop is a strict cap of three cycles. If the initial broad intent is too vague, the refinement phases may not fully assemble the required context set within the limited iterations.