Listen

Description

> Watch the video version for the live Realm Labs demos: https://www.youtube.com/live/EDXHJnwPdjQ

Agents have access to sensitive tools and data, run long multi-turn tasks, and act on their own. Saurabh Shintre, founder of Realm Labs, and Arvind Murali Mohan, who leads product there, joined Melanie and co-host Jwahir to show how Realm catches an agent drifting off course before it acts, by reading the model's intent instead of only checking what goes in and comes out.

Arvind ran the same prompt injection scenario 20 times against a three-agent shopping system and showed one run quietly adding a counterfeit product to the cart, then walked through how Realm's intent signals and policies flag and stop that path at runtime. Saurabh demoed Deep Neural Inspection, Realm's production take on mechanistic interpretability, showing a model's internal state shift in real time before it generated a single word of its response.

We get into:

- Why evals and red teaming can't fully cover a non-deterministic agent: same prompt, same model, different outcomes

- Catching refusals, warnings, and prompt injection as intent signals before the tool call runs

- Policies as a signal plus an action: block, redact, reprompt, reroute to another model, or alert

- Deriving policies from your system prompt and baselining normal agent behavior

- Why keyword filters and LLM as a judge fall short when you need to intervene in real time

- Reading model internals at the token level, including running an open-weights "digital twin" for teams on closed frontier models

- How behavior differs across model families when they refuse or push back

- False positives, thresholds, and universal, industry-specific, and business-specific policies

- A live voice agent demo with signals for user frustration and PII firing as the conversation happened

Resources:

- Realm Labs: https://www.realmlabs.ai

- RealmGuard playground: https://playground.realmlabs.ai

- See inside AI's brain (Deep Neural Inspection demo): https://dni.realmlabs.ai

- Why Agents Go Rogue: https://www.realmlabs.ai/resources/why-agents-go-rogue

- Temporal: https://temporal.io

- Temporal on GitHub: https://github.com/temporalio

- Temporal Community Slack: https://temporal.io/community

Vibe Check is Temporal's near-weekly livestream on AI, developer tooling, and durable execution. Subscribe so you catch the next one.

---

Originally streamed November 4, 2026 on Vibe Check Temporal YouTube channel and you can watch the full episode at: https://www.youtube.com/live/EDXHJnwPdjQ