> All I'm saying is it doesn't matter which one you end up using, as long as the job gets done.
> The outcome is far more important for our users than letting them choose which model they want to use.
This has been my approach with the current client.
Their developers originally had very strong pushback regarding the selection of one specific provider, but we persisted long enough for everyone to realize that not chasing perfect vendor agnosticism can actually result in some really impressive solutions. The code can easily be refactored to use a different provider, but we couldn't have gotten to where we are if we had to support all providers every step along the way.
Normalizing OAI, Anthropic & Google at every step for an enterprise chatbot that needs extremely deep custom tooling is a very cursed mission. I grant you that you can make it work perfectly fine for a generic developer toolchain. Shell and patching files is not exactly high variance between providers. I think HN is kind of blind to other use cases outside code monkey activities. I am more worried about things like PDF analysis and nested context detection in complex desktop screenshots. Can all providers meet these goals? Kinda? At the edges you have to start getting very particular with the specific provider's quirks.
At a certain point it is easier to just maintain a branch per provider once you get the overall solution close for one. Attempting a code-level interface is a complete joke of a quest. You need to modify prompts, string templates, event loops, etc. There is no clean way to do this if you are building genuinely novel agents.
bob1029 · · focus · HN ↗
> The outcome is far more important for our users than letting them choose which model they want to use.
This has been my approach with the current client.
Their developers originally had very strong pushback regarding the selection of one specific provider, but we persisted long enough for everyone to realize that not chasing perfect vendor agnosticism can actually result in some really impressive solutions. The code can easily be refactored to use a different provider, but we couldn't have gotten to where we are if we had to support all providers every step along the way.
Normalizing OAI, Anthropic & Google at every step for an enterprise chatbot that needs extremely deep custom tooling is a very cursed mission. I grant you that you can make it work perfectly fine for a generic developer toolchain. Shell and patching files is not exactly high variance between providers. I think HN is kind of blind to other use cases outside code monkey activities. I am more worried about things like PDF analysis and nested context detection in complex desktop screenshots. Can all providers meet these goals? Kinda? At the edges you have to start getting very particular with the specific provider's quirks.
At a certain point it is easier to just maintain a branch per provider once you get the overall solution close for one. Attempting a code-level interface is a complete joke of a quest. You need to modify prompts, string templates, event loops, etc. There is no clean way to do this if you are building genuinely novel agents.