Jev System One: Typed Decisions for Agent Pipelines

What it is: Jev is TypeSafe AI's "System One model," launched 2026-09-15 — a non-autoregressive, non-chat model that plugs into an agent pipeline as a fast typed-decision step rather than a generation step. Instead of prompting an LLM and parsing free text, you send it unstructured program state and get back a typed answer — Choice (categorical), Score (numeric/rubric), or Noul (yes/no probability) — from a schema you define up front. TypeSafe AI's own positioning: "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." This page reflects a five-day-old launch; treat install and API details as best-known-as-of-launch, and confirm exact SDK method names against TypeSafe's official docs before shipping.

Why it matters for agent frameworks

Agent pipelines are full of small, repeated, typed decisions hiding inside prompts — "which tool should I call," "is this output safe to return," "how urgent is this ticket" — that get answered today with a full LLM call and a parsed JSON response. Jev is pitched as a purpose-built layer for exactly that class of decision: TypeSafe AI's own framing is "a fast decision layer that classifies, scores, and routes," escalating to a frontier LLM only for the genuinely open-ended steps. Slotted into an agent loop, it can sit in front of a frontier model as a router (see our LLM routing patterns tip) or behind one as a guardrail (see guardrailing LLM outputs).

Install & configure

Official SDKs exist for Python and JavaScript. At the level we're confident in without TypeSafe's exact package documentation:

# Python
pip install typesafe-sdk

# JavaScript / Node
# Optional: compare Jev against chat models on the same questions
pip install 'system-one-adapter[openai]'

Under the hood, both SDKs are wrappers around a plain HTTP POST:

POST https://api.typesafe.ai/v1/systemone
model: jev-latest

Early access is currently waitlist-gated via console.typesafe.ai, which also hosts a playground for testing schemas before wiring them into an agent. We were not given exact SDK method names, request field names, or auth header formats beyond the endpoint and model route above — check TypeSafe's official documentation for those specifics rather than assuming a shape here.

What can I plug it into?

Any point in an agent framework where the "decision" is really a Choice, Score, or Noul question in disguise — see our Jev question types tip for worked examples. TypeSafe also publishes an open-source adapter, referenced as system-one-adapter-python, so an existing LLM in your stack can emit Jev-compatible structured outputs rather than calling Jev's API directly — useful if you want the fallback path and the fast path to share one schema definition.

How do I keep it honest?

Do not treat a schema-valid response as a correct one. TypeSafe reports a 0% structured-output error rate for Jev, meaning output always matches your schema — but on TypeSafe's own four-workflow benchmark, Jev's accuracy is 67.8%, close to GPT-5.6 Terra (67.9%) and behind Sol (74.1%) and Opus 5 (73.1%). None of these numbers are independently reproduced yet. Before routing real decisions on Jev's output, validate accuracy and calibration on your own data — see calibrated confidence scores and structured-output error rates for how.

Troubleshooting

Most early problems come from the schema rather than the model: a question whose options exceed the cardinality cap, program state passed as an image, or a confidence threshold set without checking calibration first.

• No API access yet: early access is waitlist-gated as of launch — sign up at console.typesafe.ai.
• Category list rejected: Choice schemas cap at 255 categories; split a larger taxonomy into a hierarchy of narrower Choice questions.
• Image input fails: Jev accepts text-based structured state only — images aren't supported at launch.
• Output looks "too confident": check calibration on your own data before trusting a Score or Noul threshold in an automated decision.

Author & links

Author: TypeSafe AI (official)

Homepage: typesafe.ai

Adapter repo (referenced): system-one-adapter-python

Related skills and tips

For the underlying category and model mechanics, see what is a System One model, non-autoregressive models, and RLCD explained. For agent-framework comparisons, see LangGraph and Pydantic AI for how typed-output patterns look on the LLM side of an agent stack.

← Back to Agent Frameworks