iterative-retrieval

Subagent context is refined using iterative retrieval, which applies a 4-phase loop (DISPATCH, EVALUATE, REFINE, LOOP) with relevance scoring. It progressively converges on high-quality information through up to three cycles for multi-agent workflows.

Iterative retrieval manages context size constraints by starting with broad queries and progressively refining the information given to subagents. This prevents context overload by evaluating and filtering relevance across up to three loop cycles.

Iterative retrieval works by dispatching broad queries, evaluating the returned context for relevance, refining the query based on that evaluation, and looping this process. It converges on necessary codebase context over up to three cycles.

Yes, iterative retrieval is designed for multi-agent workflows with evolving requirements. Subagents often do not know what context they need until they start working; the 4-phase loop progressively gathers and refines information to meet these dynamic needs.

Iterative retrieval is needed when spawning subagents that require codebase context, handling complex multi-agent workflows, or facing context size constraints. It solves the problem of subagents not knowing what context they need until they begin working.