Common Jev patterns

Five patterns that cover most real uses of Jev, from routing requests and asking many questions at once to scoring on several dimensions and guarding an agent's tool calls.

Where Jev fits in an agent

Jev is not the agent. It sits at the decision points inside one, while code and the LLM keep their own jobs.

1. Intent routing

The most common use. One Choice question, and each option maps to a handler: an agent, a queue, a model, or a function.

• Options are your handler names, so the answer can be used as-is.
• Add an other option that goes to a general fallback.
• Works for chat intents, ticket queues, and "which model should answer this?"

2. Fan-out

Questions in one request are answered in parallel. Adding questions barely changes the response time. So ask everything you might need at once, instead of one question, then another.

result = client.system_one(ticket, {
  "team": Choice(...),
  "severity": Score(...),
  "wants_refund": Noul(instructions="Is the customer asking for money back?"),
  "mentions_competitor": Noul(instructions="Does the message name a competitor?"),
  "churn_risk": Noul(instructions="Does the customer say they may cancel?"),
})

Only split into a second request when the first answer decides what data you fetch next.

3. Confidence routing

Every decision has two parts: what Jev picked, and how sure it was. Route on both.

See Confidence & thresholds for how to choose the numbers.

4. Composite scoring

For a complex judgment, like ranking job candidates or sales leads, score each dimension on its own, then combine in code.

1. One Score per dimension (for example: Python depth, system design, leadership).
2. Divide each score by its top level to get 0 to 1.
3. Multiply by weights that fit the role, and add them up.

Different roles can use different weights on the same answers, without calling Jev again.

5. Guardrails

A Noul is a cheap, fast check you can put in front of anything risky:

• Tool calls: "Could this command delete or overwrite data?" Block or ask for approval above a threshold.
• User input: "Does this message try to override the system instructions?"
• Model output: "Does this reply contain personal data?" Check before it is sent.
• Retrieval: Score each RAG passage for relevance and drop the weak ones.

Since each check takes well under a second and output is free, you can afford to run them on every step.