TypeSafe AI's JEV picks an action without generating tokens: what's inside

⚡ The gist in 5 seconds

• TypeSafe AI has built JEV — a network that picks an action inside code instead of composing text.
• Developers call it the first model of the System One Models class; its open-source counterpart is Laya.
• In the breakdown on Habr, Laya and Gemma 4 run on your own server in a FastAPI service.

🔍 What happened

Python developer Alexey Yakovenko wrote a lengthy breakdown (a 54-minute read) for the HOSTKEY blog about JEV — a network from TypeSafe AI. It doesn't hold conversations, doesn't write code, and doesn't generate arbitrary text. Its job is to pick one option from a pre-known list: trigger RAG, go to search, call an API. The full text is available on Habr.

The mechanics are simple — and that's exactly what makes them uncomfortable for large models. A regular LLM produces its answer token by token even on a task like this. Structured Outputs, native Tool Calling, and JSON Schema-based generation make the format predictable, but they don't change the autoregressive principle. Hence the author's question: why use a generative model where there's nothing to generate. TypeSafe AI answers with its own term — System One Models, a nod to Kahneman: System 1 is fast and intuitive, System 2 is slow and deliberate.

The practical part is about something else. We take Laya, an open-source counterpart of JEV, put Gemma 4 alongside it, and build a news agent. Laya reads the material and decides whether the topic is interesting. Then a large LLM analyzes the text, prepares an article, and publishes it to a blog. All of this lives in a FastAPI service: a database, a background task worker, a custom Agent Harness, and a web interface for watching the process. No LangGraph or CrewAI — the logic is pure Python.

🧠 Our take

Agents have matured, and almost every step they take is a tool choice. Each such choice is paid for in tokens and latency. Our assessment: moving routing into a separate small model makes sense — classification is cheaper than generation. But the "System One" name itself is a nice repackaging of an old idea. Routers and classifiers have long lived in agent pipelines; they just weren't called a new class.

What does TypeSafe AI's bet change? If JEV and its counterparts match LLM accuracy when choosing among ten options, agent architecture will be turned upside down. The large model stays for the hard stuff, the small one takes over the routine. This is a direct conversation about your wallet: fewer tokens, fewer seconds of waiting, less dependence on someone else's API. For now it's a promise, not a measurement — and that's where I'd pump the brakes.

Who wins? Those building agents on their own hardware: open-source Laya removes vendor lock-in. Who loses? It looks like the "decision models" niche won't go entirely to niche teams. Large LLM platforms would sooner build such a router into their API than give away a slice of the market. We believe that in a year this will be a subscription feature, not a standalone product.

⚠️ What we don't know

The source contains no model sizes, no latency figures, and no comparison against Tool Calling on the same set of tasks. Nothing is said about what license Laya ships under or how much hardware it needs. Everything we know about JEV comes from TypeSafe AI's own claims as relayed by the tutorial's author. You'll have to verify it yourself: run the router and the LLM on your own tasks and compare accuracy, time, and cost.

Prediction: within a quarter, at least one major LLM platform will add a separate cheap decision router to its API. The niche has already been described, and open-source counterparts lower the barrier to entry.

Disclaimer: one source — the breakdown on the HOSTKEY blog; claims about JEV come from TypeSafe AI and are not confirmed by independent tests.

Source: Habr / HOSTKEY