Heyo guys. Just saw the new baml site and was completely blown away. Seems like a lot has changed for the better. I had a couple of questions regarding the new language... sorry if its repetitive... feel free to simply point me to external docs. I've been working a lot making specialized or domain-specific harness at work (fancy way of saying building stuff on top of pi core) and was wondering if i could potentially replace pi core with baml. Curious to know if baml will: - be able to use some non traditional vendor apis (so I can use the codex sub as an api provider instead of the openai sdk). More specifically how easy would it be for me to add support for bespoke llm providers? - the docs implies the api will be multi modal enough to build voice agents... so I'm assuming it will support text to speech and speech to text? What if we have bespoke endpoints to consume? - what about provider specific quirks like interleaved thinking and stuff - prompt caching. Im sure that will be covered but will I get to control things... for exanple if i want to inject additional tools which will bust the cache, can I gracefully recover from it either by recaching on demand or make a new warm cache preemptively. - Support provider specific tool calling mechanics... i love bamls structured responses but I dont want to have to choose between llm native tool calling or baml structured output. Cant i have both? Most importantly... will baml support some kind of repl environment or code hot reloading (like elixir). So I can dynamically extend the harness for self improvement loops?

🙂

This is soo cool. Looking forward

if you're into it we do onboardings weekly: https://www.boundaryml.com/eap

we do have plans to 'load' old traces and cache the results of functions to be able to tweak a workflow, but no ETA on that

Registered

I went through this page: https://boundaryml.com/explore But I'm afraid it didnt quite click with me. I mostly asked the question cause a couple of my friends recently bought up the topic of BAML and i wanted to see stuff for myself. Here's some feedback (I'm keeping it blunt in the spirit of transparency) My perfective going in: "Cool, a language for agents... so how does this help me build agentic systems better?" 1. Part 1 is mostly about its features as a programming language. Out of experience i know thats awesome... but none of those are problems for me today. Typescript with the effect library gives me a "good-enough" layer. 2. Part 2 is where things start getting interesting. 1st and 2nd point are good to know. But 3 onwards is what excites me. It shows that baml can be treated almost as an interpreted language now. And given my past experience knowing how baml functions work i know how awesome this can be. 3. The workflow explorer is what really got my attention. Cause it reminded me the kind of low level control baml could potentially have on llm responses. The whole workflow conversation that i had with Vaibhav on my stream. 4. Part 5 was the most important piece (the thing i actually was trying to find since baml is supposed to be a language for agents after all). But it doesn't go into any details I wanted to see. I was really hoping to get some insights on how this helps me conpared to the other tools i am evaluating (like PI, claude agent sdk, even flue framework).