iterative-retrieval

Subagent context discovery improves by applying an iterative retrieval loop that progressively refines candidate files through a four-phase cycle: DISPATCH, EVALUATE, REFINE, and LOOP. This staged approach reduces wasted work and closes context gaps in complex workflows.

The best way to refine relevant files is using automated query refinement based on relevance signals and discovered terminology. The system quantitatively scores file relevance and dynamically updates search queries across multiple cycles.

Iterative retrieval handles oversearch by enforcing a strict maximum three-cycle limit. This guardrail prevents excessive searching while still allowing the EVALUATE and REFINE phases to adequately score relevance and refine queries.

You need iterative retrieval when spawning subagents in codebases or complex workflows where context is revealed only after initial tasks. It enables staged discovery of relevant files and terminology when upfront context is opaque or delayed.

Yes, iterative retrieval scales to large codebases and complex workflows. By progressively returning a refined set of relevant files and terms through its loop mechanism, it enables smarter subagent orchestration without overwhelming the context window.