Use case · Gate · Score · Next Action

Use Jev for Agent Guardrails

Add fast structured checks around tool-using agents without asking a frontier model to deliberate on every step.

Input

Agent state + proposed action

Structured state sent to the decision layer.

Decision

allow · review · retry · stop

Allow, escalate, retry, stop, or ask the user?

fastPolicy checks
typedOutcomes
loopAgent runtime
humanEscalation option
How it works

Turn the workflow into decisions.

Capture proposed tool call and context.

Evaluate it against a clear policy.

Allow low-risk actions.

Escalate uncertain or high-risk cases.

Log results for evaluation and threshold tuning.

Try this pattern

Allow, escalate, retry, stop, or ask the user?

state: { source: "...", context: "..." } question: "Allow, escalate, retry, stop, or ask the user?" pattern: "Gate · Score · Next Action"

When not to use Jev

If the task requires open-ended writing, deep synthesis, or novel code generation, keep a generative model in that step. Use Jev around the decision points.