Using the Jev model in CowAgent

[zhayujie]

Jev is a model built for fast, structured judgments — classification, tagging, scoring, and decisions in milliseconds, at very low cost. Getting it running in CowAgent takes just 3 steps:

1️⃣ Sign up at console.typesafe.ai and grab your API Key
2️⃣ Install the Jev skill with one command: /skill install typesafe-ai
(or send CowAgent the skill link and it installs automatically: https://raw.githubusercontent.com/typesafe-ai/skills/main/skills/typesafe-ai/SKILL.md)
3️⃣ Describe what you need in natural language, and let CowAgent build it for you with Jev

The screenshot below is a ticket-triage demo built with CowAgent + Jev: paste in a raw customer ticket and a single call returns multi-dimensional results — department, emotion, priority, churn risk, and more — each with classification and scores.

Anything that's a "judgment you can make at a glance" and runs at high volume is a great fit for Jev.

[aniruddhaadak80]

Short answer, and it may not be the one you were hoping for: Jev can't drive the chat turn itself, but it can sit alongside the chat model as a cheap decision layer.

Why not as the model

models/custom_provider.py documents api_base as "required, must be OpenAI-compatible", and every provider resolves through the OpenAI-compatible surface — a chat-completions request/response cycle.

System One is not that. Its API reference is an evaluation endpoint: you send state plus typed questions and get typed answers back with calibrated probabilities and a confidence value. The primitives are Choice, Score, and Noul. Their own primer puts it plainly — TypeSafe trains "decision models with calibrated probabilities instead of optimizing for generated text" — and Confidence exists precisely because a System One model is meant to decide, not to reply.

So pointing custom_api_base at it and setting model: jev-latest would fail at the first request, since there's no text generation to satisfy a chat turn. Same reason it won't work as a custom_providers entry either — the requirement is on the endpoint, not on the config shape.

What it is actually good for here

The part I'd genuinely suggest: Jev is cheap and fast enough to sit in front of the big model, where a full LLM call would be wasteful. Their docs have end-to-end recipes for exactly the shapes an agent loop already has:

• Guardrails — one Noul per inbound message, thresholded in code to pass / review / block. Cheaper than a completion and you get a probability instead of a vibe.
• Confidence-gated routing — let Jev answer, and let confidence decide whether to act on the answer or escalate. This is the pattern that makes it safe to auto-apply a decision.
• Intent routing — a Choice over your tool set, dispatching to a deterministic branch, a specialist model, or a human.

That last one fits CowAgent's structure better than the others, because models/bot_factory.py already routes on bot_type — a Choice that picks a provider and returns a calibrated confidence is the same decision with an attached number.

If I've misread the question

I checked the System One docs rather than assuming, and I only see the evaluation endpoint — no OpenAI-compatible surface. If TypeSafe ships one that I didn't find, or if you had something narrower in mind (a specific sub-task rather than the main turn), say so and I'll dig further.

Sources: models/custom_provider.py at c94f3bba; docs.typesafe.ai/llms.txt and the Guardrails / Confidence / Intent-routing pages under it.