iterative-retrieval

You refine codebase context for subagents by iteratively retrieving information across up to three cycles. This workflow uses a four-phase loop—DISPATCH, EVALUATE, REFINE, and LOOP—to progressively narrow down relevant files for tasks like debugging or feature development.

The context problem in multi-agent workflows occurs when subagents do not know what information they need until they start working. This issue is solved by using an iterative retrieval loop that automatically evaluates and refines queries to fill information gaps as the task progresses.

Iterative retrieval decides to stop searching when it finds three high-relevance files and fills the context gaps. The workflow evaluates relevance scores during the EVALUATE phase and automatically stops early once this specific condition is met.

Yes, you can use iterative context refinement for codebase exploration, debugging, and feature development. It is specifically designed for scenarios where the required context is unclear until work begins, progressively guiding deterministic retrieval across up to three cycles.

A limitation of this iterative retrieval loop is that it is constrained to a maximum of three cycles. Additionally, the stop condition is strictly tied to finding three high-relevance files, which may not be sufficient for extremely large or complex codebase explorations requiring deeper context.