Is Jev an AI agent? Three ways to put a model in your software
By Christophe Paka · Facts checked on 23 October 2026
Jev is TypeSafe's AI model that answers narrow questions about your data, picking from a list of answers you wrote. A reader asked someone to explain Jev like they are not a dev. Another asked if Jev could replace the model in a coding agent. Both questions hit the same wall: where does the model sit in your software?
The short answer: no, Jev is not an agent. So, is Jev an AI agent or something else? The figure below answers with one test: who owns the control flow, the order in which steps run.
How to read this figure
Start with traditional software. The left panel is a decision tree of if checks and steps. Code owns every step, so the pieces are reliable and fit together. But as the figure says, a rule cannot read intent in free text, such as the words of a customer ticket.
Follow the loop of an LLM agent. An LLM agent is an AI that picks its own next step in a loop. The middle panel shows steps 1 to 8, then a question mark, then a dashed line back to the start. TypeSafe's design guide argues that each turn of the loop is one more chance for the model to pick a wrong step. It gives no data for this; read it as a design argument.
Compare the shape of the third panel. The right panel, AI powered software, has the same tree shape as the first one. That is the point of the drawing: code still owns the flow. The model does not choose what runs next.
Find the two Jev nodes. Two of the decision points are Jev nodes. The first sorts a ticket into billing or technical. The second answers yes or no. Each one returns a typed answer: one of the fixed answers you listed, never free text. So your code knows every path in advance.
Read the rule under the panel. Jev answers narrow typed questions only where common sense is needed. TypeSafe's line: System One is its model for building AI powered software, not agents. It does not write code or choose its own next action.
Key takeaways
- Keep the control flow in your code, and call Jev only at the points that need judgment.
- Swap a fragile rule, such as a keyword list, for one narrow Jev question with fixed answers.
- Use an LLM, not Jev, when you need text written or an agent's next step chosen.
Try this: find the Jev nodes in your own flow
You need a pen, a sheet of paper and one workflow your team runs, such as ticket triage. No code yet.
- Draw the workflow as a tree of checks and steps, like the left panel.
- Circle each check that has to read free text, such as a ticket or an email.
- For each circle, write one question with a fixed list of answers, such as billing or technical.
- Try that question with a real example in TypeSafe's Playground. It is a page in TypeSafe's web console, at console.typesafe.ai/playground: you log in, paste your example as the state, add the question and see the answer. Runs there may be billed, so try a few, not hundreds.
Expected result: A tree where code owns every arrow, with one or two Jev nodes at the points that need common sense. To turn one node into code, read the post on the first call, step by step.
Questions readers ask
Can Jev replace the model in my coding agent?
No. TypeSafe's docs say no model setting turns a coding agent into a Jev powered agent. Jev does not write code or choose its own next action. What helps instead is TypeSafe's agent skill, a plugin that teaches a coding agent to write Jev code.
Can someone explain Jev like I'm not a dev?
Picture your workflow as a flowchart. Most boxes are plain steps. A few are questions that need common sense, such as which team should handle this ticket. Jev answers only those questions, from a list you wrote. Your flowchart then decides what happens next.
What is the difference between an AI agent and AI powered software?
An agent picks its own next step, again and again, in a loop. In AI powered software, your code runs the steps in the order you wrote. It calls a model only for narrow decisions, then acts on the typed answer. Jev is built for the second way.
When is a trained classifier better than Jev?
A trained classifier is a small model fitted to your own labels. The book's rule of thumb: it wins once the question is stable and you hold thousands of labels, or when it must run offline. Jev needs no labels to start, and you change the question without retraining. No published test sets the break even point, so run both on the same labeled set. The post on the right tool for the job draws the full test.
Sources: How to build with System One (TypeSafe docs); System One models (TypeSafe docs).
This figure is from Jev Made Crystal Clear, chapter 2.
The chapter builds the idea step by step, with recipes you can follow, a Try it to check yourself, and every source.
Get the book
