Every agent loop you have built in the last year does the same expensive thing at every fork: it asks a language model to write a sentence about what should happen next, then scrapes a decision out of that sentence. Route the ticket. Pick the tool. Escalate or not. The prose is discarded. The branch is kept. You paid for the prose.
That pattern was never a law of nature. It was a side effect of the only models that worked. Last week a company founded by one of the people who made chat models follow instructions shipped a model that will not write a sentence at all.
A function call that returns probabilities
TypeSafe AI came out of stealth on 15 September 2026 with Jev, the first public member of a class they call System One models. Simon Willison and others prefer a blunter name: decision models. The interface is simple enough to sound like a joke. You send a state (text, a record, a blob of program context) and a list of typed questions. You get back floats.
- Choice - pick among options you defined; probability mass over each option plus a confidence.
- Score - place the state on a rubric you described in words; a continuous position along that scale.
- Noul - a yes/no framed as a Bernoulli; a single number in [0, 1]. The CEO confirmed on Hacker News that the name is short for Bernoulli.
Multiple questions ride on one state and are evaluated in parallel. Adding questions barely moves latency. There is no token stream to wait for, because there is no generation. TypeSafe prices input at $0.042 per million tokens and charges nothing for output. Vendor claims put end-to-end latency in the 70-500 ms band and workflow speedups on the order of 40×-200× versus frontier chat models on classification-shaped work. Homepage numbers go higher still (roughly 194× faster and 445× cheaper on their own workflow suite). Those figures are TypeSafe's; treat them as vendor claims until third parties publish matched harnesses. Even at a discount, the economic shape is the story: you stop buying discarded prose.
LangChain already ships a TypeSafeClassifier and middleware for model routing and "auto mode" tool guardrails. Browser agents have started letting Jev pick the DOM operation while a small LLM only wakes up when actual text must be typed. The split is not theoretical. The stack is already growing a second path.
The inventor walked the other way
Diogo Almeida, TypeSafe's founder, helped build the instruction-following methods that became the research spine of ChatGPT. In the launch essay he says the quiet part: models have been superhuman at chat for years, and automation still lagged. Chat, he decided, was not the missing organ. The missing organ was a model trained to make calibrated decisions that software can consume without a human reading them.
The training slogan is the tell. The industry spent half a decade on RLHF - reinforcement learning from human feedback, optimized for what raters prefer to read. TypeSafe trains with RLCD - reinforcement learning for calibrated decisions. Same family of tools. Opposite target. The person who helped teach models to talk is now shipping a model that is not allowed to talk, on purpose.
Daniel Kahneman's System 1 / System 2 split was always a metaphor about human minds. TypeSafe turned the metaphor into a product category. System One here means: fast, parallel, no narration, branchable. The chat model stays for the work that still needs words - drafts, plans, code, explanations. The decision model takes the forks that were never literature.
"Can't hallucinate" cuts both ways
TypeSafe's marketing line is sharp: because Jev does not generate strings, it "can't hallucinate." Schema matching is guaranteed. The output type is fixed before the call. That is a real property. It is also a trap if you stop reading there.
Willison's reaction is the useful counterweight. LLMs were already black boxes that could at least invent a justification. Jev returns only a float. If it marks a message as spam, or a candidate as low-fit, or a city as "good," there is no sentence to audit. Bias does not disappear when you delete the prose; it hides inside the number. Willison's Bay Area city experiment - Cupertino on top, East Palo Alto on the bottom for a vague "good city" score - is the kind of result that should make any hiring or credit pipeline flinch. Jaggedness docs already warn that Jev is weak on numbers, dates, and adversarial content. A model that cannot write also cannot warn you in English that it is guessing.
So the safety claim and the opacity claim are the same claim. No text means no confabulation. No text also means no confession. Teams that bolt Jev onto agent guardrails get cheaper, faster vetoes. They also inherit a classifier they cannot cross-examine. The correct response is not to reject the class. It is to treat decision models the way serious shops treat ranking models: eval sets, slice metrics, and a human path when confidence is low. TypeSafe exposes confidence for that reason. Confidence is only useful if someone writes the threshold in code.
What this does to the agent bill
Look at a typical coding or ops agent. A large fraction of model calls are not "write the patch." They are:
- Is this tool call dangerous?
- Which of five tools fits?
- Should we escalate to the expensive model?
- Is the retrieved passage relevant?
- Did the last step succeed enough to continue?
Those are System One questions. For two years the industry answered them with System Two machinery - frontier chat models, structured-output prompts, JSON parse retries, and a prayer that the model would not monologue. Structured outputs were a bandage on a text generator. Jev is a different organ wearing the same problem statement.
That does not retire Claude, Astra, or Fable. Open-ended reasoning still needs a language model. It does change the default architecture diagram. The honest diagram now has two boxes: a writer and a voter. The writer is expensive and legible. The voter is cheap and mute. Harnesses that keep paying writer prices for voter work will look quaint by next quarter, the way hand-rolled intent classifiers looked after the first good embedding routers landed.
The open-weight reaction time is already a signal. Within days of launch, community projects stood up Jev-shaped interfaces on small open models, left-pad jokes, a 2048 player, and a deliberately terrible chat bot that picks the next symbol with Choice questions. When a new shape is real, people rebuild the shape before they trust the vendor. That is happening.
The half you were overpaying for
The discovery is not "another startup shipped a classifier." Classifiers are old. The discovery is that the agent boom forced a confession: a huge share of "intelligence" in production agents was never generative. It was routing, scoring, and gating - work we only gave to chat models because chat models were what we had, and because RLHF had made them good enough to fake a switch.
Almeida's arc makes the confession personal. The research line that taught machines to converse also taught the industry to overuse conversation. System One models are the correction. They will be misused on decisions that need explanation. They will also delete a tax every agent has been paying in silence: the cost of words nobody was going to read.
The agent loop did not get smarter this week. It got a second brain. That brain does not speak. If your harness still asks a novelist which door to open, you are not being careful. You are being nostalgic.
Research note (2026-09-23): X/Twitter search via xurl was unavailable (API CreditsDepleted / HTTP 402). This brief used primary web sources only. Key sources: TypeSafe launch post (typesafe.ai/blog/introducing-system-one-models-and-jev, 2026-09-15); TypeSafe docs (docs.typesafe.ai); Simon Willison, "Jev introduces a new shape of LLM" (2026-09-21); LangChain "Building a Harness with Jev" (2026-09-17); systemonemodels.org independent hub. No X posts cited. publish_path=supabase+neon
