Webinars / Past sessions
Jev for Evaluating Agent Actions: Dynamic Rules w/ Deterministic Enforcement (13 Min Live Event)
Recording coming soon
This session has wrapped and the recording is still being processed. Check back shortly — it lands here as soon as it's published.
Browse other sessionsWhat you'll learn
- Using model judgment to evaluate state without making the model the control
- Typed questions, probabilities, and policy decisions made in application code
- Demo: checking a proposed Slack escalation against a customer's support email
About this session
Not every control decision is a clean yes or no. Whether a proposed action fits the current task, matches user intent, or looks like normal behavior for this agent is often a judgment call, and that’s exactly the kind of check a model is well suited to make. The mistake is letting that judgment also be the enforcement. Jev, a decision model built for exactly this role, evaluates state against typed questions and returns a probability, but that answer only ever informs a policy decision made in application code before it reaches a tool.
This session covers how to bring model judgment into agent controls without letting the model become the control itself, so enforcement stays deterministic even when the questions being asked aren’t. We’ll demonstrate in a customer support context by checking a proposed Slack escalation against a customer’s support email before allowing it to send.
Hosted by
Thierry Damiba
Member of Technical Staff
Thierry Damiba is a Member of Technical Staff at Arcade.dev, focused on what it takes to run MCP and AI agents in production inside the enterprise. He writes playbooks from the field on agents, GTM engineering, and what happens when machines start doing the work. He’s the founder of Scale Intelligence and was previously a developer advocate at Qdrant.
