“I think it should be like 90% of this is like we wrote deterministic old school code and we only call the LLM for very specific parts of this where we have to interpret language and make decisions. That's the only thing the LLM is doing.”
Startup to Last · Bootstrapped & Indie SaaS · November 2026
Context
Tyler King on his new AI product bet. Agents burn tokens at every step, so for business software he wants mostly ordinary code, with the model called only where language has to be interpreted.
Transcript
like, even if you schedule a task, my impression of what's happening is it is taking the text prompt that whatever you give and giving it to the LLM fresh every single time. That's you might get a different result every single time.
…like the basic. Yeah. What you can't, what they also have is the concept of artifacts, which are kind of coded like bodies of work that are static from a container perspective, but dynamic within that container. And then you can have the scheduled task do stuff within that container. Yeah, I've
done a bit of that. Again, I'm being vague because like this is still blurry and we're still making giving it shape. I think there's a lot more to do there where instead of 10% of stuff is predefined in an artifact and 90% is LLMs burning tokens on the fly. I think like most SAS, this isn't true for all LLM use, but for like SaaS use cases, I think it should be like 90% of this is like we wrote deterministic old school code and we only call the LLM for very specific parts of this where we have to interpret language and make decisions. That's the only thing the LLM is doing.
Yep.
I think there's an opportunity there.
I agree. I 100% agree. It's a challenge every day for me is getting to that.
More from Startup to Last
