Matei Zaharia Introduces Open-Source Omnigent for Contextual Agent Security

Databricks co-founder and CTO Matei Zaharia released Omnigent, an open-source agent meta-harness designed to address the overly simplistic security controls of current coding agents. The framework can be layered on top of existing agents like Claude Code and Codex, introducing "contextual policies" to track session states and enable more flexible, dynamic security management.

Core Mechanics: State Tracking and Intent-Driven Security

Zaharia points out that traditional agent security is limited to static action lists like allow/deny/ask. Omnigent's contextual policies introduce state tracking: implementing "intent-driven" security by setting base permissions according to the user's initial prompt, while also recording real-time metrics during the session to dynamically adjust subsequent control policies.

Key Applications: Risk Control and Cost Management

By tracking the session state, Omnigent handles complex security scenarios. For instance, it logs whether an agent has read confidential documents or untrusted external content. If confidential docs were accessed, it requires human approval before sending emails; if untrusted content was touched, it limits prompt injection risks. Additionally, the framework calculates risk scores based on actions like web browsing or software downloads, and supports real-time cost accumulation (e.g., pausing to ask the user every $10 spent) for granular cost and risk management. Omnigent comes with built-in practical policies and allows users to write custom ones based on official documentation.

2026-07-08 ~ 2026-07-08 · 6 related posts

Primary sources