7/7 My take: Jev matters less as new general intelligence than as a missing control-plane primitive. A generative model proposes; Jev makes bounded semantic branches; deterministic code enforces the non-negotiables.
1/7 Why is everyone talking about Jev? It became one of the week's biggest AI launches despite being unable to write a sentence. That limitation is exactly the point. 🧵
6/7 Jev cannot emit an invalid schema, but it can return a wrong valid answer. Adversarial state can move risk judgments. Use it as one guardrail layer, never as the security boundary.
6/6 My current principle: deterministic code protects non-negotiable constraints; agents handle non-enumerable change. Generated code and actions must stay within explicit permission, testing, review, budget, and rollback boundaries. How do you draw this line?
1/6 After building an agent harness, I realized that agentic software isn’t just a traditional workflow with an LLM inserted into one step. It may be a fundamentally different software architecture. 🧵
5/6 The central question is no longer only “How do we make code open for extension?” It is also: What must remain deterministic? What may be delegated? How are agent actions tested, reviewed, observed, and rolled back?