Jev's layered architecture: LLM writes, Jev decides, code acts
The author argues that yes/no decisions in agent loops, such as routing, scoring and approvals, should not be generated by large models but handled by Jev as structured decisions.
We’re using billion-dollar AI models to answer yes/no questions That’s the problem Jev Engineering is trying to fix Most agent loops constantly need to decide things like: - which worker goes next - whether a source is relevant - whether a tool call should be approved - whether to retry or escalate None of that really needs long-form generation That’s the idea behind Jev Engineering: LLM writes Jev decides Code acts LLMs handle open-ended generation Jev handles routing, scoring and structured decisions Code handles deterministic execution My take Jev is interesting not because it replaces LLMs It’s interesting because it removes work that probably never belonged to an LLM in the first place And I think this is where AI systems are heading: not one giant model doing everything but specialized layers doing the job they’re actually good at The next leap in AI might not be a smarter model It might be a smarter architecture