iterative-retrieval

To build context for subagents, use iterative retrieval to progressively narrow down relevant files and patterns. This approach uses a four-phase loop—DISPATCH, EVALUATE, REFINE, LOOP—across up to three cycles to surface high-relevance files for downstream tasks.

The best way to narrow codebase context exceeding token limits is progressive context collection. By iteratively evaluating and refining search criteria, the loop minimizes data transfer and produces a curated set of pertinent files without overwhelming the context window.

Iterative retrieval refines context by looping through four phases: DISPATCH to search candidate files, EVALUATE to assess relevance, REFINE to update criteria, and LOOP to repeat. This cycle runs up to three times to progressively surface pertinent files.

Yes, you can use progressive context collection for code exploration in dynamic projects. The iterative loop works for multi-agent workflows and code exploration by building a reusable strategy to discover required context with minimal data transfer across evolving codebases.

A limitation of iterative retrieval is that it caps at three cycles to manage context size. When refining context for subagents, the loop may not surface every pertinent file if the codebase is exceptionally large or highly fragmented across unrelated modules.