new | show | ask | jobs Github

ekropotin 15 hours ago

Dynamic code generation for calling APIs, not sure what is a fancy term for this approach.

▲

gzalo 14 hours ago | parent | next [-]

Something like https://github.com/huggingface/smolagents

Needs a sandbox, otherwise blindly executing generated code is not acceptable

▲

ianbutler 11 hours ago | parent | prev | next [-]

https://www.anthropic.com/engineering/advanced-tool-use#:~:t...

Anthropic themselves support this style of tool calling with code first party now too.

	▲	ekropotin 11 hours ago \| parent [-]
		Yup, that’s I’ve been taking about.

▲

inerte 12 hours ago | parent | prev | next [-]

Cloudflare published this article which I guess can be relevant https://blog.cloudflare.com/code-mode/

▲

willahmad 15 hours ago | parent | prev [-]

this assumes generated code is always correct and does exactly what's needed.

▲

ekropotin 13 hours ago | parent [-]

Same for MCP - there is always a chance an agent will mess up the tool use.

This kind of LLM’s non-determinism is something you have to live with. And it’s the reason why I personally think the whole agents thing is way over-hyped - who need systems that only work 2 times out of 3, lol.

▲

anon84873628 12 hours ago | parent [-]

The fraction is a lot higher than 2/3 and tool calls are how you give it useful determinism.

▲

ekropotin 11 hours ago | parent [-]

Even if each agent has 95% reliability, with just 5 agents in the loop the whole thing is just 77% reliable.

	▲	anon84873628 4 hours ago \| parent [-]
		Well fortunately that's not what actually happens in practice.