"I just saw the news about the Grok Bot source code leak, and one point really stood out: they designed 'send message' itself as a tool call. That's exactly the same direction we explored before — not replying directly with a raw assistant message, but using BAML for structured output. Same underlying philosophy. But there's a known pitfall: once you drop in a weaker model, or one that's been heavily RL-tuned, it tends to bypass the pattern and respond with a plain assistant message, breaking the format and causing parsing errors. We've been there, and it looks like Grok is running into the same issue. They'll probably need RL or fine-tuning to patch it. Still, in hindsight, this approach was pretty forward-looking. The controllability is unmatched — you keep full ownership of the output. The trade-off, though, is that it puts a heavy burden on the base model's capability. Otherwise, you end up doing what Grok Bot is doing now: layering on extra prompts just to enforce constraints."

its all just code all the way down! i think more and more folks are going to want pure code representations of things for control. (mostly cause people will always want some sort of UI that isn't plain terminal)