iterative-retrieval

To progressively refine subagent context in a large codebase, use an iterative retrieval loop with four phases: DISPATCH, EVALUATE, REFINE, and LOOP. This cycle scores relevance and dynamically updates search criteria to assemble the necessary files for subagents.

The best way to assemble relevant files for multi-agent code exploration is using iterative retrieval. It scores context relevance, identifies missing information, and reformulates queries until sufficient context is assembled for the subagents.

You need iterative retrieval for subagent context management when spawning multiple subagents for bug fixes or feature development in large projects where they cannot predict the required codebase context upfront.

The four-phase retrieval loop works by dispatching subagents, evaluating relevance scores, refining search criteria to close context gaps, and looping until the correct codebase information is pruned and assembled.

Yes, you can adjust the retrieval cycles. The iterative retrieval loop features adjustable cycles, allowing you to control how many times the context is dynamically refined and queried before assembling the final relevant files.

Subagents start with limited context during codebase exploration because they cannot predict the necessary files upfront. Iterative retrieval solves this by progressively narrowing and refining the information load through dynamic query refinement.