Hey guys, please lmk if i'm posting this in the wrong channel (don’t really use discord) but is anybody here interested in programatic tool calling for agents? So writing code w/ the tools acting like api's

I’m interested for sure! I feel like it all just boils down to remote code execution and “programmatic tool calling” is just the feature organization for it

i some know services like https://www.daytona.io/ or backend frameworks like deno that make spinning up sandboxes that lets you run code pretty straight-forward

Haven’t tried Daytona but I am using Modal Sandboxes for a project

not exactly sure rn, i guess initially i would like to see how this would stack up against a traditional agent with the same tasks, how reliable it is, the number of tokens, number of steps and etc

Fair alright. True, probably benefits seen more at scale or higher complexity of task

ya i doubt this being useful for smaller agents w/ only a few tools, but the more tools you have w/ each request taking more and more tool calls is where i think this would be useful

There was a blog from vercel recently where they stripped down an internally used agent to just having bash access, and that improved quality and speed both

oh nice! would you mind sharing the link

i remember watching a similar example of that for the github mcp, like bash vs tools where the bash way more efficient because the llms where good (and trained) at writing the cli commands rather than tool calling

Ah have not watched this one yet, but i do know about the idea. I’ve been working on a repro of this tool called byterover, which is a CLI memory management system for coding agents. They recently pivoted from mcp to cli so agents could invoke it better. The reproduction: https://github.com/Sanjan611/bitranger

Interesting! So this is like a shared & updated memory for your codebase for every ai agent to reference / update via the CLI?

Correct. ByteRover makes it a cloud-based solution so anyone across any ide or coding agent can consume it, but for my purpose i was just making it locally available

Cool man! Been taking a look, do you have a demo of this in action or any potential use cases?

Not yet it’s a pretty new endeavor cause I’m interested in the idea of “memory”. Will share a demo once it’s a bit better!

I am using this SDK to implement memory and sandbox, although it is not yet ready for production use https://turso.tech/blog/agentfs

Cool! what problem were you solving with this, and what’s been most useful so far?

bash