[ Use case · Agents ]
Can Jev make decisions inside an AI agent?
Yes, for the choice steps. Agents spend much of their time picking the next action from a known list. Jev can make that pick in tens to hundreds of milliseconds while a language model handles planning and writing. Reported browser-agent runs cost fractions of a cent per task.
Is it a fit?
Use Jev when
- Your agent repeatedly chooses from a known set of actions or elements.
- Latency matters: games, browsers, real-time systems.
- You already use a language model for planning and want to cut its calls.
Skip it when
- The step requires writing, coding or multi-step reasoning. Jev can't plan.
- The set of possible actions isn't known in advance.
The questions
One request per step, 2 questions. Jev answers them in parallel.
- Choice
Which element should the agent click next to make progress on the goal?
Options: 1 · 2 · 3 · 4
- Noul (yes/no)
Has the goal been fully completed on this page?
Options: Yes · No
Getting started
[ 1 · Test it in the playground ]
No code needed. Paste a real step into TypeSafe's playground and ask: “Which element should the agent click next to make progress on the goal?” with the options 1, 2, 3, 4. You'll see the answer and its probability.
Open the playground (opens in a new tab)[ 2 · Use an existing tool ]
We haven't found a ready-made tool for this workload yet. These directories track what's been built on Jev:
Made with Jev (opens in a new tab)Jev.Store (opens in a new tab)JevHunt (opens in a new tab)
[ 3 · Integrate it ]
To run it on every step automatically, call Jev from the system where your steps live. The developer details below include the full request.
What people have reported
Browser agent: flight search
self-reported1 task for under 1¢, 7 seconds
Gregor Zunic, co-founder of Browser Use · Made with Jev (opens in a new tab)
Minecraft gameplay agent
self-reported1 session (8m 43s) for 97¢
Ronak Malde · Made with Jev (opens in a new tab)
Self-reported figures are the author's own; we haven't reproduced them. More on the costs page.
Common questions
- Can Jev replace the LLM in my agent?
- No. It can replace the calls where the agent picks from a list. Planning, reasoning and writing still need a language model.
[ For developers ]Request body, wiring and pitfalls
The request
One request per step with all 2 questions batched over the same state, so Jev reads the state once. Keys below are the names you'll see in the response.
{
"model": "jev-latest",
"state": "Goal: Book the cheapest nonstop SFO→JFK on Oct 12.\nPage: Search results. Visible elements: [1] Sort by price, [2] Nonstop filter, [3] First result $412 (1 stop), [4] Next page",
"questions": {
"next_action": {
"type": "choice",
"instructions": "Which element should the agent click next to make progress on the goal?",
"criteria": {
"1": "Sort by price",
"2": "Nonstop filter",
"3": "First result",
"4": "Next page"
}
},
"goal_done": {
"type": "noul",
"instructions": "Has the goal been fully completed on this page?"
}
}
}How to wire it up
- 1.Let your planner (a text model) decide the goal and sub-goal.
- 2.At each step, send the goal plus the list of visible elements to Jev as a Choice.
- 3.Fall back to the text model when confidence is low or no option fits.
Watch out for
- Large pages mean large state. Send the element list, not the raw HTML.
- Watch the 32k-token limit on state plus the longest question.
New to writing Jev questions? Read the question design guide.
Related use cases
Can Jev score and route inbound leads?
Rank inbound leads by fit and intent before a human looks at them.
Can Jev triage an inbox?
Sort mail into categories, rate urgency and flag what needs a reply.
Can Jev route support tickets to the right team?
Send each ticket to the right queue and flag frustrated customers.