@theramblingfool@theo So it’s his responsibility to carefully explain the nuances when you misunderstand how things work? Why?
He’s said since the beginning of t3 code that it’s not a harness.
If you ask any decent model it will tell you t3 code is not a harness.
So it seems like willful ignorance
@sheherenow_ 😔 first try npx t3 triage and see if that helps you troubleshoot at all?
I did have some weird quirks and trouble getting it to run exactly how I wanted a few months ago but once I did it’s been pretty great since. I use it almost exclusively from my phone now.
@maria_rcks@theo Make it customizable and provide a bunch of options out of the box. Your addition of the art gave t3 code a really nice vibe.
I copied it for a bb plugin so I could get the same vibe
@snortydogg@sheherenow_ I will have to tinker with this later then! I wasn’t fully sure when I first read through the README.
Have you found the models effective at writing workflows that invert the control and only call back to the agents for judgement?
@sheherenow_@snortydogg I’m almost convinced to retry one of my experiments but with a hacked up version of effect+typescript. If it was Friday night I might be doing it already
@snortydogg@sheherenow_ They should have access to the same context, the same tools. They should be partners fully aware of each other and when and how to hand off
@snortydogg@sheherenow_ I’ve played with a few experiments but it really needs more time than I can probably dedicate.
I want there to be symmetry between an agent and code. The agent can write and call code. That code should be able to write and call agent prompts, as part of the same session.
@rupertmanfredi@getdiffer Sure. Still early and experimental, but I think it's simple enough to be useful (for me, at least!)
https://t.co/kWzSHVgDjK
100% on the shared building blocks. Need better primitives for agents that are at a higher abstraction level. Cant wait to see what is done in this space.
@dexhorthy Do you think adding capabilities to the harness and a language like typescript with libraries is enough to get closer to that goal? Or do we need a language designed for agents and models trained on the language?