A Jev AI agent is an agent that hands its small decisions to Jev — TypeSafe AI's decisions-only model — so the loop picks, routes and clicks in under half a second while a writing model only gets called when something actually needs writing. That one split changes how your whole agent feels. Most of what an agent does all day is not writing. It is choosing: which button, which tool, which model, which lane. Jev answers exactly those questions, fast and almost free, and it flat-out refuses to do anything else.
📺 Watch: NEV Jev AI Voice Browser Agent is WILD!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I have spent the past week building with this and filming two videos on it, and the same pattern kept paying off: pull the decisions out of your expensive model, hand them to a dedicated decider, and the whole agent gets faster, cheaper and far more predictable.
What a Jev AI Agent Actually Is
Jev is the System One model from TypeSafe AI, founded by Diogo Almeida, who previously worked at OpenAI. Early access opened on 15 September 2026, and I wrote a plain-English breakdown in what Jev AI actually is.
The constraint is the product: Jev never writes text. You send a question plus a set of options, and it returns a pick and a confidence score. Nothing else. It speaks three question types — Choice, which picks one option from a list capped at 255 choices; Score, which rates something; and Noul, a yes-or-no probability. Answers land in 70 to 500 milliseconds. Input costs $0.042 per million tokens and output is free, because a pick is barely any output at all.
So a Jev AI agent is a division of labour. The decider handles every fork in the road: which element to click, which agent takes this task, which model gets this request, whether the user is even finished talking. The writer — Claude, GPT, whichever model you already like — only wakes up when a human needs to read a sentence. That moves the expensive model from every step of the loop to almost none of them.
The Voice Browser Agent, Piece by Piece
The clearest working example is the one I filmed in "Jev AI Voice Browser Agent is WILD!" — an open-source MIT project by Moritz Kremb that lets you drive a real browser with your voice. The repo is jev-voice-browser on GitHub, at github.com/moritzkremb/jev-voice-browser, and I went through the code in my repo breakdown.
The agent is three plain parts. The browser's built-in speech-to-text turns your voice into words. Jev decides what those words mean for the page. Playwright then moves a real Chromium window — clicking, typing, scrolling, switching tabs.
The clever part is not what decides. It is when the decisions happen. The agent re-asks Jev on every new word you speak. Each ask lands well under half a second, and the moment the next word arrives, the stale ask gets cancelled and a fresh one goes out. The result feels borderline rude: the page is often loading as the last word leaves your mouth. You never wait for the agent to think, because it was thinking the whole time you were talking.
And each ask is not one question. Every request bundles about ten questions at once: what action to take, which element to act on, which site this is about, whether the sentence is finished, whether you are even talking to the agent, how far to scroll, and which of your words belong in the search box. In my runs, each request carried nine questions and came back in about 400 milliseconds. One round trip, nine decisions.
This decider-first browser pattern is exactly what we build together inside AI Profit Boardroom — 3,000+ members shipping agents, at $69/mo locked in (normally $110). Prefer to map it onto your own business first? Book a free strategy session and I will walk your stack with you.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
My Speed Test: One Decider Against Three Writers
I ran my own small test while filming. Eight voice commands, with Jev answering nine questions per command. Median response: 445 milliseconds. Score: 8 of 8 correct. Then I gave three writing models a far easier job — one question per command instead of nine — on the same eight commands.
| Model | Questions per command | Median response | Correct |
|---|---|---|---|
| Jev | Nine | 445 ms | 8 of 8 |
| Claude Haiku | One | 1.6 s | 7 of 8 |
| Gemini 3.8 Flash | One | 2.2 s | 7 of 8 |
| GPT 5.6 Luna | One | 2.5 s | 7 of 8 |
My caveat, because it matters: this was eight commands, not a proper benchmark, and the other models were not tuned for this job. Treat it as a smell test. But the shape of it — the decider clearing nine questions before the writers could finish one — matches everything else I saw this week. When I want a real head-to-head, I run it through Goldie Bench.
📺 Watch: Hermes AI Agents Just Went Portable
Why This Agent Cannot Hallucinate Its Way Into Trouble
Hallucination is the reason most people never let a browser agent near anything real. The voice agent's answer is structural — four guardrails enforced by code, not by prompts.
- It can only click what exists. Before each decision, the app snapshots up to 100 clickable elements and gives each a short label. Jev answers with a label. It cannot click something that is not there, because things that do not exist never make it into the options.
- It cannot put words in your mouth. The code cuts your sentence into pieces and Jev only points at one. The code then copies your words into the box letter for letter. A pointing model cannot invent text.
- It cannot wander off. The site list is owned by the code. Jev just picks from it.
- It knows when to ask you. When confidence drops under 0.45, the app pins numbered badges on the page and asks which one you meant. Your reply never even goes to the model.
Notice the shape of every guardrail: the decision surface is a closed list, so the worst case is a wrong pick from real options — never an invented action. You design the safety in, instead of begging for it in a prompt.
Jev Agent Patterns Beyond the Browser
The browser is the flashy demo, but the decider pattern generalises. I covered ten builds in "Jev AI: 10 INSANE Use Cases", and the full list lives in my use cases post. Three stand out for agent builders.
The self-assigning task board. Cards drop in, and the options Jev sees are whichever agents are available right now. Jev picks a lane for each card, and anything it is unsure about goes to a "you" lane instead. A confidence line decides what moves on its own — and that single line is what lets the board keep running while you walk away.
The model router. LangChain shipped a ready-made piece for this — langchain-typesafe — where you describe each model's strengths in plain English and Jev routes every request to the right one. It keeps a live counter of what you actually spent versus what sending everything to the expensive model would have cost. Watching that router counter climb is the fastest way to convince anyone the decider pays for itself.
The browser-loop rebuild. A browser-agent team I cite in the video rebuilt their loop so Jev picks every action and a small writing model only fills text boxes. Their numbers, not mine: flights found in 7 seconds for under half a cent, browser commands down from 192 to 101, and task time down 25 percent. Same job, half the chatter.
Wiring the Decider Into Your Own Stack
You do not need to rebuild anything to start. Get a key at console.typesafe.ai, find the noisiest fork in your current agent loop — usually routing or element selection — and hand that one decision to Jev. My setup walkthrough is in how to use Jev.
In my own stack, the decider is becoming a layer. I am adding it to my Agent OS as Jev matures, sitting between intake and the writing models — the architecture is in my Agent OS guide. The Score and Noul question types also slot straight into the critic seat of my gauntlet loop: a cheap, sub-second gate that screens work before an expensive model touches it.
Then set your confidence lines deliberately. Every Jev answer arrives with a confidence score, so pick a number per decision type: above the line, the agent moves alone; below it, the item queues for you. The voice browser agent uses 0.45 and badges. Your task board might run higher. Either way, autonomy becomes a dial you set, not a vibe you hope for.
Jev AI Agent Questions, Answered
Does a Jev AI agent replace my writing model?
No — the split is the point. Jev makes the picks: routing, clicking, scoring, yes-or-no calls. Your writing model still writes anything a human will read. Jev never produces text, so it could not take the writer's seat even if you wanted it to.
What does Jev cost inside an agent loop?
Input is $0.042 per million tokens and output is free. Decisions are tiny, so even the voice agent's ask-on-every-word approach stays cheap, and the LangChain router's live counter shows the saving against an everything-to-the-big-model setup in real time.
What happens when Jev is not confident?
Every pick ships with a confidence score, and your code decides what to do with it. The voice browser agent pins numbered badges under 0.45 and asks you. The task board routes unsure cards to a "you" lane. That confidence handling is yours to design, which is exactly why these agents are safe to leave running.
How many options can one Jev question hold?
A Choice question caps at 255 options. In practice you narrow first — the voice agent snapshots up to 100 clickable elements, well inside the cap — and you can bundle around ten questions into one request, which is how it packs nine decisions into a single 400-millisecond round trip.
Start Building the Decider Layer
The agents that feel like magic right now are not running bigger models. They are running fewer writers and faster deciders, and Jev is the first model built purely for that seat. Clone the voice browser repo, move one fork of your loop onto a decider, and you will feel the difference in a weekend.
If you want to build it alongside people already shipping this, join us in AI Profit Boardroom — the builds, the prompts and the member agents are all inside. And if you want a personal plan for where the decider fits in your business, book your free strategy session and we will map it together.
