Skip to content
Jev + Agentium: decisions, tools, and evaluations.See how
Agentium

Jev, inside your agent application

Give Jev a place in the whole workflow.

Use Jev for the decision. Use Agentium to connect it to the context, agents, and evaluation around it.

Jev is a decision model built and operated by TypeSafe AI. Agentium integrates it through the @typesafe-ai/sdk; Agentium does not build, host, or sell Jev.

Start with JevSee three ways to use it

Labels. Probabilities. Rubric scores. Results your code can act on.

Choose where judgment belongs.

  1. 01

    A decision agent.

    Send the relevant facts and define the question. Use the returned decision in your own routing and application logic.

    Use Jev as a model
    decision-agent.ts
    import { Agent, choice, jev } from "@agentium/core";

const inboxRouter = new Agent({
  name: "inbox-router",
  model: jev("jev-latest"),
});

const decision = await inboxRouter.run(
  "Please send a copy of the invoice for my subscription.",
  {
    questions: {
      destination: choice("Which team should receive this request?", {
        accounts: "Invoices and subscription payments",
        engineering: "Product defects and API failures",
        general: "Other requests",
      }),
    },
  },
);

console.log(JSON.parse(decision.text).destination.choice);
    Install
    npm install @agentium/core @typesafe-ai/sdk
    Server-side env
    TYPESAFE_API_KEY
    Note
    Validate parsed output before using it for an action.
    Open guide

  2. 02

    A tool for your chat agent.

    Keep the conversation with your chat model. Add Jev tools when the agent needs a judgment during the task.

    Add the Jev toolkit
    service-desk.ts
    import { Agent, openai } from "@agentium/core";
import { JevToolkit } from "@agentium/core/toolkits";

const serviceDesk = new Agent({
  name: "service-desk",
  model: openai("gpt-6-astra"),
  tools: [...new JevToolkit().getTools()],
  instructions:
    "Ask a Jev tool to classify the request, then explain the next step.",
});

console.log((await serviceDesk.run("I cannot download my invoice.")).text);
    Install
    npm install @agentium/core openai @typesafe-ai/sdk
    Server-side env
    OPENAI_API_KEY, TYPESAFE_API_KEY
    Note
    A production policy may need explicit validation that the tool was called.
    Open guide

  3. 03

    A judge in your evaluations.

    Run a response through criteria you define, then turn the result into an explicit pass rule.

    Build a Jev scorer

    The runnable recipe lives in the docs. Open the Jev evaluation guide.

A decision is useful when the rest of the system can use it.

You define the questions, the thresholds, and what happens next.

  • Bring context

    Pass the information the decision depends on.

  • Connect the next step

    Use the result in your agent tools or application logic.

  • Keep human review

    Use your own thresholds and approval policies for consequential actions.

  • Check the behavior

    Evaluate responses against criteria relevant to your product.

Technical details

Jev handles typed decisions. Use a chat model when the application needs a written response.

In the normal questions API, it returns a value from 0 to 1. Your application decides the threshold or action.

Yes. Wrap the Jev call in a custom scorer. The current documentation does not use a dedicated jevJudge helper.

Install Agentium core and the TypeSafe SDK. Set your TypeSafe API key on the server. Add the eval package if you are building an evaluation.

Put a decision to work.

Start with one question. Connect the answer to the next step.

Open Jev docsExplore Agentium