iterative-retrieval

Context retrieval for subagents is optimized by iteratively searching, evaluating relevance, and narrowing results up to three cycles to progressively find the right context without overloading the system.

Subagents fail with incorrect context because they receive irrelevant noise or insufficient data, causing wasted tokens; a structured iterative search loop with relevance scoring progressively finds the right context.

Query refinement and exclusion of low-relevance files are achieved by ranking files with a 0–1 relevance score, selecting high-value context, and refining search keywords across multiple bounded cycles.

Iterative search cycles should be used for agent orchestration when initial context needs discovery, such as focused code reviews or targeted bug fixes, and token budgets must be respected.

The maximum number of search loops for context retrieval is three cycles, or the loop stops immediately upon discovery of sufficient high-relevance files scoring 0.7 or above.

Iterative context retrieval works for focused code reviews with strict token budgets by applying relevance scoring and query refinement to exclude low-relevance paths and return only high-value context.