The solution to this is to have a separate prompt (on a tiny shitty model) run to determine first if any tools should be run at all, and if so, further drill down to what categories or individual tools should be presented to the tool-using model.

Dead simple queries either don't use the tool-capable model at all or are just outright answered by the prompt evaluator.

This is how AgentV3N solves this problem while also bringing down the cost of allowing *every* query potentially use a tool.

https://venturebeat.com/ai/langchain-shows-ai-agents-arent-human-level-yet-because-theyre-overwhelmed-by-tools/

Reply to this note

Please Login to reply.

Discussion

Also, less tools can be more. The very *first* tool I added was Powershell.

From there, I was able to ask for very very advanced things and even the use of a few web APIs for things like weather and traffic. Why? Because it was just using Invoke-WebRequest against public APIs it knew about.

Just with access to a shell you have a very capable assistant. But that's also dangerous as fuck which is why a good 80% of this codebase is permissions and authorizations