TypeSafe Opens a Playbook So Coding Agents Make Smarter Decisions With Jev
TypeSafe released an open agent-skills pack that teaches Claude Code, Cursor, and Codex how to call its Jev model for typed, calibrated decisions.
Takeaways
- TypeSafe published an open agent-skills repo for its System One API, MIT licensed.
- Installs as a Claude Code plugin or via
npx skills add typesafe-ai/skillsfor other agents. - Wraps Jev, a model returning typed decisions with calibrated probabilities instead of free-form text.
- Uses a new training algorithm called Reinforcement Learning for Calibrated Decisions (RLCD).
- Skill teaches agents to keep policy in code and route uncertain cases to human review.
- Targets ticket routing, moderation gates, tool selection, and search pruning in agent loops.
TypeSafe has released an open-source skills repository for coding agents including Claude Code, Cursor, and Codex. Its core artifact is a single SKILL.md file that teaches agents how to build workflows around Jev, TypeSafe’s model for typed decisions with calibrated probabilities.
Installing the skill adds integration guidance to an agent’s context. Jev remains available through TypeSafe’s System One API, which requires the usual credentials and client configuration.
One file defines the pattern
TypeSafe uses the term “System One Models” for models designed to make bounded decisions inside software. Jev accepts structured questions, selects from allowed values, and returns a probability distribution that application code can evaluate.
TypeSafe says Jev uses a new architecture, sampler, and training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The skill presents these model calls as small, reusable judgments that can be composed into larger workflows. Application code retains control of orchestration, retries, thresholds, business rules, and final actions.
Probabilities become policy inputs
A machine-consumable decision needs a stable type and a useful uncertainty signal. Jev returns an allowed value alongside probabilities for the available choices, giving software a direct input for automation and escalation rules.