@NirantK Why don’t we do make this real, when your APIs are ready give me like a beta access along with your eval set queries. Let me run with what I am building and share time, cost and response we can compare…
That works ?
1. Having a large library of tools ain’t tool explosion, infact it’s inevitable if you have a thriving ecosystem of agents. Some tools might be shared.
2. Having a lot of tools hot in model context is a problem.
3. And you don’t have to. (Low key you don’t even need MCP.)
(If you squint, tools (APIs in your case) are just one abstraction of capabilities (more dev friendly) we can add to an agent and skills (say a MD file) are a different abstraction of capabilities(more layperson friendly).
Openclaw and clawhub fx, the hub might have 1000s of skills. Would you say they have a skills explosion problem just for having a large hub. They don’t inject it all into the system prompt and keep them hot. (Albeit will admit in this case end user needs to install skills and that’s bad for your usecase).
You don’t need it to be a hot list and you don’t need the users to install it. That’s the best middle ground :)
To be continued ..
@NirantK You can’t be mad at douchebags :), take it easy let’s ship that well sought after harness. Albeit frontier LLMs are not the only kind of models we can ship.
Tool Search Tool as a meta tool for searching tools exists. Anthropic released it for their ecosystem before PTC programmatic tool call. Dynamic tool composition is not a leap if you can search tools reliably for a query. Since TST was locked to one ecosystem @andersonbcdefg wrote an implementation on dynamic composition across providers and wrote a blog about (if my memory serves).
“Nobody knows what that means” seems like a projection of yours and that possibly gets you going :)
I would argue tools are akin to ingredients and skills are akin to recipes. As frontier moves LLMs will cook with their own recipes i.e. we can leap to Dynamic Tool Composition with retrieval confined to Tool searches (on a large tool library which external tool makers will bring) . So skills and skill search will be subsumed at worst or reserved only for non-frontier models at best.
Progressive disclosure and skill is a big part of agent harness engineering.
Effective retrieval over skills is going to be big, might even become "web scale".
"Under this paradigm, we propose Skill Retrieval Augmented Agents (SR-Agents), which dynamically retrieve and use relevant skills from large-scale skill corpora to expand their problem-solving capabilities"
https://t.co/dIETJBAUQg