Jev and your own decision model in one AI agent

Every agent makes small decisions before each step: which tool to call, whether an input is safe, whether it has enough to answer. Most teams send all of them to an LLM.

This demo shows a support agent in Domino that sends them to decision models instead and calls the LLM only to write the reply and take the cases the decision models aren’t sure about. Jev (TypeSafe) picks the next tool and diagnoses the failure, and the Domino Decision Model, a small model trained on our own prompt injection examples and served as a Domino endpoint, screens tool output before the agent reads it.

DDM is a proof of concept, built to show the architecture works end to end, not to compete with Jev. Every call goes through LLM Gateway 2.0, so each model runs under the same budget, guardrails, and record, and swapping in a new decision model like CLM-8B is a gateway change with no edits to the agent code.

00:00 — Why do agentic AI workflows overuse the LLM?
00:36 — What is a decision model?
00:57 — Which model should handle which decision?
01:46 — How does an agent screen for prompt injection?
02:10 — What happens when a decision model isn’t sure?
02:24 — How does agentic AI governance work in an LLM gateway?
03:00 — How do you train your own decision model?