How We Improved Agentic Search

An analysis of 1,983 public coding-agent checkpoints and 202,142 tool calls from the Entire CLI repository found that search-related actions accounted for nearly half of agent tool use, spanning file retrieval, shell-based searching, and direct content search. Researchers compared raw ripgrep, the faster indexed search server fff, and pgr, a ranked agent-oriented search tool, using public benchmarks based on real agent prompts and traces. Although fff substantially reduced individual search latency, it produced only modest end-to-end improvements because tool execution represented a very small share of total runtime relative to model inference and planning. pgr’s definition-first, path-aware ranking improved retrieval quality, particularly for first queries and implementation tasks, increasing the likelihood that a relevant file appeared first while returning less output than raw ripgrep. Broader end-to-end tests showed modest reductions in search calls, cost, and wall-clock time for pgr, but high task variability prevented consistent improvements in total tool calls or overall efficiency. The study concludes that result ranking and presentation appear more influential than raw search speed for helping agents identify useful code early, while noting that larger, more diverse evaluations across models and semantic retrieval approaches are needed.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.