SkillSeek halves agent-skill retrieval spend

SkillSeek, an open-source two-stage skill retriever, reached observed parity with an LLM-mediated retrieval loop on the 89-task SkillsBench benchmark while cutting per-trial spend from $51.30 to $27.54, according to the authors’ arXiv paper. It runs on CPU with zero in-loop LLM tokens, but the result is one benchmark-and-harness configuration—not a universal replacement; Octobench’s same-model harness gap is a reminder to test skill routing inside your own agent system before paying for agentic search.