iterative-retrieval

Subagent file discovery is handled by progressively refining context through a four-phase loop of dispatching, evaluating, refining, and looping. This iterative retrieval process identifies relevant files, learns codebase terminology, and scores file relevance within a maximum of three cycles.

Multi-agent workflows waste cycles when subagents lack initial repository context and must blindly search for code patterns. A progressive context refinement loop solves this by iteratively re-querying the codebase, learning naming conventions, and narrowing focus to high-relevance files before execution.

Iterative retrieval is a context refinement technique for multi-agent systems. It uses a four-phase loop—dispatch, evaluate, refine, and loop—to progressively query the codebase, score file relevance, and collect high-relevance context within a maximum of three cycles.

Context explosion is mitigated by iteratively refining subagent queries and focusing strictly on high-relevance files. By scoring relevance and limiting the discovery loop to three cycles, the system learns project structure and terminology without ingesting unnecessary repository data.

Iterative retrieval is well-suited for debugging multi-agent tasks. It progressively refines context to help subagents learn repository structure, naming conventions, and key code patterns, allowing them to efficiently complete assignments within three refinement cycles.

The primary limitation is the maximum three-cycle constraint on the refinement loop. If the codebase terminology or file structure is highly fragmented, subagents may not fully discover all relevant context before the loop terminates, potentially leaving gaps in their assignment context.