收录解读
This paper studies retrieval inside agent loops rather than retrieval as an isolated offline component, comparing grep and vector retrieval across different agent harnesses and result-delivery modes.
The key finding is that retrieval performance depends strongly on orchestration: whether results are inline, file-based, or mediated by a specific CLI/harness can change which retrieval strategy wins.
The work is valuable because it turns a common engineering assumption into an empirical systems question: agentic search quality is a joint property of retriever, harness, tool interface, and context consumption path.
For agent engineering, it provides a useful evaluation caution and a strong baseline argument: simple lexical search can dominate when the harness exposes evidence in a form the model can actually use.