Skip to main content
Your agents make dozens of small judgment calls a day: is this draft on-brand, does this reply answer the question, which task should come first. Usually they judge their own work. Jev, from TypeSafe.ai, gives them a second opinion: a purpose-built evaluation model that returns typed, calibrated decisions in milliseconds, at a fraction of the cost of a full model call.

​
What it does

Jev answers questions about a piece of content with three answer types:
  • Yes/no with confidence (“Was the ticket resolved quickly?”)
  • One choice from a list (“Which of these four categories fits?”)
  • A score on a scale (“Rate this draft 1 to 10”)
Every answer ships with a calibrated probability, so an agent can act on confident answers and flag uncertain ones for you instead of guessing.

​
What agents use it for

A few patterns that work well:
  • Self-review before shipping. An agent drafts an email, then asks Jev whether the tone is professional and the ask is clear before sending.
  • Routing. An inbound request comes in; Jev picks which agent or workflow should handle it.
  • Quality gates in routines. A scheduled run scores each batch of output and only surfaces the ones above threshold.
  • Cheap checks in loops. When an agent would otherwise call a full model to grade something, Jev does it faster and for a fraction of the cost.

​
How to use it

Ask your agent in plain language:
Under the hood, agents call it with the platform beta jev command. The full syntax, flags, and question format are in the beta commands reference.
Jev is a beta integration: it works today, and the command surface may change as it matures. Authentication runs through the platform’s built-in proxy, so there’s no API key to set up.

Jev command reference

Flags, question types, and examples for platform beta jev.