iterative-retrieval

Subagent context retrieval improves by progressively narrowing the search space and pruning irrelevant files through a four-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP, running up to three cycles to converge on high-quality context.

The subagent context problem occurs when agents start with limited context and struggle to identify relevant files, terminology, and patterns across large codebases, leading to incomplete understanding and poor task execution.

Iterative refinement works by dispatching broad queries, evaluating results for relevance, filtering out low-relevance files, and looping the refined search up to three cycles to enrich context and avoid information overload.

Use iterative retrieval when subagents operate in large codebases where a single search pass returns too many irrelevant files, requiring progressive relevance-based filtering to identify accurate terminology and patterns.

Yes, the iterative retrieval loop is capped at three cycles with relevance-based filtering at each stage to prevent context overload and ensure subagents receive only high-quality, relevant file information.

Agents fail when starting with limited context and no refinement mechanism, making it difficult to identify relevant files and patterns; an iterative retrieval loop progressively narrows the search space to solve this.