iterative-retrieval

You can improve subagent context by applying an iterative retrieval pattern with a four-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. This structured process progressively narrows retrieved materials to align with the subagent's specific task.

The best way to refine retrieved context is through iterative retrieval that dynamically prunes irrelevant files and detects context gaps. It enforces a deterministic cycle, evaluating relevance and refining search criteria up to three times to balance depth and latency.

Start iterative retrieval by issuing a broad initial query to the subagent. The system then evaluates the relevance of the retrieved materials, prunes irrelevant files, and refines search criteria through a maximum of three deterministic cycles.

A maximum of three refinement cycles is enforced for subagent context retrieval. This deterministic limit ensures a balance between achieving deep context alignment and maintaining acceptable execution latency in multi-agent workflows.

Multi-agent workflows handle missing context by using context-aware pruning to dynamically exclude irrelevant files and identify gaps. The system then refines search criteria in an iterative loop, progressively guiding subagents toward relevant files and concepts.

You should avoid using an iterative retrieval pattern when the subagent already possesses sufficient context, or when the latency introduced by up to three refinement cycles outweighs the benefits of dynamically pruning irrelevant files and detecting context gaps.