iterative-retrieval

Iterative retrieval solves context discovery for subagents by progressively refining search results through a four-stage loop. It uses relevance scoring and dynamic keyword extraction to shrink large codebases into high-confidence context artifacts.

Multi-agent context discovery is best managed by iteratively refining retrieved information to match subagent needs. This approach applies pattern-based queries and evaluation feedback to expand or prune search terms, ensuring progressively accurate context provisioning.

Iterative retrieval refines context by executing a four-stage loop: DISPATCH, EVALUATE, REFINE, and LOOP. It evaluates initial results, extracts dynamic keywords, adjusts pattern-based queries, and repeats until cycle limits are met or high-confidence context is produced.

Yes, iterative retrieval optimizes token usage by staging context extraction from large repositories. It dynamically prunes irrelevant search terms and results through evaluation feedback, ensuring subagents receive only the most relevant, high-confidence context artifacts.

You need iterative retrieval when subagents require progressively accurate context or when initial context is too large or incomplete. It is essential for multi-agent workflows handling complex codebases where single-pass information retrieval fails to provide relevant patterns.

The main limitation of an iterative retrieval loop is the enforced cycle limit, which stops refinement after a set number of iterations. If pattern-based queries and dynamic keyword extraction fail to converge, the resulting context artifacts may lack full relevance.