AI Agents Have 230,000 Skills — Cheap New Search Finds the Right One

AI agents (software that can actually do tasks for you, like booking, filing, or researching) rely on prepackaged instructions called "skills." Think of them as recipes an AI can pull off the shelf: one for summarizing contracts, another for cleaning up spreadsheets. Anthropic popularized this idea, and open-source collections have exploded past 230,000 recipes. The surprising problem isn't writing these skills anymore — it's finding the right one, fast, out of a quarter-million options.

The current fix is to let the AI wander the shelves itself, rewriting its search terms and second-guessing candidates over and over. It works, but it racks up costs every single time a task runs, like paying a personal shopper by the minute. A team of researchers instead borrowed the same technology behind ordinary web search: a fast first pass that narrows 230,000 skills down to a shortlist, then a slower, careful second pass that picks the winner. They published the method, SkillSeek, as open source, meaning anyone can use it for free.

On an 89-task test set, the cheap search matched the expensive AI-driven approach. The simplest version, plain keyword search, actually beat the fancy method in three of four setups; a small extra model handled the fourth. Crucially, the bill dropped roughly in half — from $51.30 to $27.54 per run, within fifty cents of using no skills at all.

The catch: this was tested on one benchmark with one setup, and the authors admit some tasks are too messy for rigid search, where an AI that thinks through the problem still wins. So this isn't "AI search beats AI thinking" — it's "use the cheap tool first." For anyone paying for AI products, cheaper plumbing usually means cheaper subscriptions down the road.