Tool Recommendation preprint reports lower costs for multi-step AI agents

A preprint reports that Tool Recommendation, a deployed system for giving AI agents access to roughly 8,200 tools, reduced bills by up to 40% in paired workflows without changing task completion. The system exposes three tools through a fixed interface instead of loading the full catalog into an agent's context. It executes calls under the user's credentials and checks service consent and tool permissions. Its first recommendation included every required tool in 93% of Codex sessions. Against a fixed list of 100 tools, input tokens fell by 17% with Codex and 40% with Claude Code. Bills fell in all 14 runs, limiting the scope of the cost evidence.