iterative-retrieval

Subagent context is improved by running a four-stage loop (DISPATCH, EVALUATE, REFINE, LOOP) that iteratively retrieves and scores relevant files, patterns, and terms across up to three cycles to return high-value artifacts.

The best way to filter low-signal results is using relevance scoring and gap identification to evaluate retrieved artifacts, progressively excluding low-value results while surfacing high-signal information for subagent execution.

Iterative retrieval works by dispatching context queries, evaluating the relevance scores of returned artifacts, refining the search gaps, and looping up to three times to discover cross-repo knowledge and surface relevant files.

You need staged retrieval when subagents enter new tasks without sufficient context, requiring progressive context refinement through multiple cycles to discover necessary files, patterns, and terms for effective task execution.

Yes, iterative retrieval specifically targets multi-agent workflows where subagents start with incomplete information, using a structured four-stage loop to progressively refine context and surface high-value artifacts for task execution.

The main limitation is that the four-stage loop runs for a maximum of three cycles, bounding the depth of progressive context refinement and gap identification available for discovering relevant files and patterns.