0:00 / 0:00

AI Agent Policies: How to Control What Agents Can Do

Name: AI Agent Policies: How to Control What Agents Can Do
Uploaded: 2026-09-02T20:03:43.111Z
Duration: 1 min 52 s
Channel: Databricks
Description: This video is part of a series on agentic AI and demonstrates how Omnigent, an open-source meta-harness, uses contextual policies to govern AI agents. It covers risk score and cost budget policies that can block actions like sending emails or require human-in-the-loop approval.

Databricks2 September 2026Watch on YouTube

Part of series

Ep. 10 · Omnigent: AI Agent Orkestratie

Databricks introduceert Omnigent, het open-source meta-framework dat meerdere AI-agenten combineert in één gecontroleerde, interoperabele workflow.

View the series

Description

Welcome to the Agentic AI Explained series. This is video 2 of 10, covering the key building blocks behind modern AI agent systems. This video covers contextual policies in Omnigent, the open source meta-harness for AI agents. An Omnigent policy tracks session state, meaning what the agent has already read, done, and spent in the current session, and uses that state to decide whether the next action should proceed. 🔗 Learn more in the blog: https://www.databricks.com/blog/contextual-policies-omnigent-using-session-state-better-govern-ai-agents 🔗 GitHub repo for this series: https://github.com/viktoriasemaan/agentic-ai-explained-labs The demo covers 2 of the built-in policies: ▪️ Risk score: each action adds to a running session score. While the score stays low the agent works uninterrupted. Once it crosses the threshold, actions like sending an email return ASK and wait for a human to sign off. ▪️ Cost budget: the policy tracks session spend against a soft checkpoint and a hard limit. At the checkpoint it pauses and asks whether to continue. At the hard limit it blocks further calls to the expensive model so the session can continue on a cheaper one. Because Omnigent is a meta-harness, the same policies apply to any agent it wraps, including Claude Code, Codex, and custom agents.

What you'll learn

You learn how Omnigent, an open-source meta-harness, uses contextual policies to govern AI agents.
You see how risk score policy works: each action adds to a session score and crossing the threshold triggers human sign-off, such as when sending emails.
You learn how cost budget policy tracks session spend, pausing at a checkpoint or switching to a cheaper model at the hard limit.
You discover that the same policies apply to any agent Omnigent wraps, including Claude Code, Codex, and custom agents.

Frequently asked questions

What is an Omnigent policy and how does it work?

An Omnigent policy tracks session state, meaning what the agent has already read, done, and spent in the current session. Based on that state it decides whether the next action should proceed.

Which two built-in policies does the demo cover?

The demo covers the risk score policy, which blocks actions and asks for human approval once a threshold is crossed, and the cost budget policy, which tracks session spend against a soft checkpoint and a hard limit.

What happens at the hard limit of the cost budget policy?

At the hard limit the policy blocks further calls to the expensive model so the session can continue on a cheaper one.

On which agents do the Omnigent policies apply?

Because Omnigent is a meta-harness, the same policies apply to any agent it wraps, including Claude Code, Codex, and custom agents.