iterative-retrieval

Subagent context improves by applying an iterative retrieval loop that progressively narrows contextual data through four phases: DISPATCH, EVALUATE, REFINE, and LOOP. This reduces wasted effort and context gaps during delegation.

Identifying context gaps during architectural analysis is best handled by iterative retrieval, which detects missing terms and files to guide targeted searches and converge on high-value context.

To start an iterative retrieval loop for code search, provide a topic and initial keywords. Iteratively dispatch searches, evaluate results, refine criteria, and loop until stopping criteria are met or enough high-quality context is gathered.

Iterative retrieval for subagent delegation is used when the context needed evolves as work progresses, such as during research tasks, exploration of unfamiliar codebases, or architectural analysis.

The four-phase workflow refines contextual data by dispatching searches, evaluating results for gaps, refining search criteria with targeted file reading, and looping until explicit stopping criteria are met.

Subagents lack necessary context because exploratory tasks require evolving information. Iterative retrieval solves this by progressively refining contextual data through an explicit loop, reducing wasted effort and context gaps.