iterative-retrieval

To refine context for subagents in a large codebase, use iterative retrieval to progressively narrow down relevant files. This approach enforces a 4-phase loop to filter low-relevance files and produce a concise set of high-relevance context for multi-agent workflows.

The best way to gather relevant files for codebase exploration with limited context windows is iterative retrieval. It applies explicit relevance scoring and dynamic query refinement across a max of three cycles to extract only the most high-relevance files for the agent.

The iterative retrieval pattern works for multi-agent workflows by executing a 4-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. It evaluates subagent context, scores file relevance, and dynamically expands search terms to narrow results over a maximum of three cycles.

You need iterative retrieval for subagent context when subagents must operate with uncertain or expanding context, such as during bug triage or feature development. It is specifically designed for scenarios where context windows are limited and upfront file relevance is unknown.

Yes, context-aware filtering helps with bug triage in multi-agent systems by excluding low-relevance files and expanding search terms based on evaluation results. This ensures subagents receive a refined, high-relevance set of files for accurate triage.

The limitation of using a 3-cycle iterative retrieval loop is that it enforces a strict maximum of three cycles to produce a concise set of high-relevance files. This caps the depth of context refinement for complex codebase exploration or feature development tasks.