iterative-retrieval

Iterative retrieval solves context management for multi-agent workflows by progressively refining information across cycles, using relevance scoring to evaluate candidate files on a 0-1 scale for just-in-time context delivery.

The best way to manage context for subagents analyzing a large codebase is iterative retrieval, which progressively narrows context across up to three cycles using a 4-phase loop and relevance scoring to deliver high-relevance context files.

To gather high-relevance context files, initialize a retrieval task and run iterations up to three cycles, using the 4-phase loop (DISPATCH, EVALUATE, REFINE, LOOP) to progressively narrow context and evaluate candidate files.

Yes, iterative retrieval works for bug fixes and new feature work where context is uncertain, applying relevance scoring on a 0-1 scale to evaluate candidate files and progressively refine the information provided to subagents.

A key limitation of iterative retrieval for context management is its cap of up to three cycles; if the 4-phase loop does not converge on high-relevance candidate files via relevance scoring within this limit, context refinement stops.