I'm curious what BAML team thinks about this? I read a few Reddit threads of people basically realizing the same thing I did: Durable Execution Engines is probably just the tried and true SWE pattern for distributed systems that will actually make an agentic system deployable in production. Sure, we can run LangGraph locally for cool demos, as can we use whatever BAML team is cooking up for orchestration, or even that github project "PocketFlow", but the problem of all of these products is that they are not designed for distributed computing. The usual junior-dev-wannabe-ai-engineer journey seems to be (source: myself and several reddit threads I uncovered): - Discover LangChain, LangGraph etc. "WOW ChatGPT can DO stuff!?" moment - (Unless you're late to the game, then you have this moment with openclaw lmao rip you) - Maybe discover something like BAML and the significance of structured outputs (12 factor agents etc.) - Makes a cool prototype that runs agent loops in-memory - (and maybe some side-effects in a DB) - Gets customers and actually needs to scale and deploy AI system to prod... - "Oh shit!!" Running everything in-memory does NOT scale, RIP all my orchestration code and hours learning LangGraph, BamlGraph... - "Wait, this seems to just be SWE? Fk, wish I wasn't a soy junior dev, SWE is hard..." - "Ok, how about a task queue system... oh yeah, Redis + BullMQ!" - Wait but... side effects in DB is not the same as function inputs... Maybe I just store function inputs as JSON dumps along with the domain-specific data in my DB? - Oh... That's what all the marketing speech was talking about on the landing page of - Well, now I've graduated into a real engineer, yay...