Omnigent: Open-Source Meta-Harness for AI Agents | Matei Zaharia
Youssef Mrini and Quentin Ambard sit down with Matei Zaharia to introduce Omnigent, Databricks’ open-source meta-harness for orchestrating, controlling, and collaborating on AI agents. Omnigent sits above the agents you already use. It wraps them with a shared layer for composition, governance, and real-time collaboration, instead of rebuilding those capabilities for every agent. In this conversation: • Compose and control multiple agents on a single task • Switch agents mid-session without losing history or context • Orchestrate work with Polly and route tasks with intelligent routing • Apply contextual policies for risk scoring, spend limits, and human approval • Run agents in cloud sandboxes you can share safely • Why Databricks open-sourced the server and runner under Apache License 2.0 Learn more: https://omnigent.ai GitHub: https://github.com/omnigent-ai/omnigent Blog: https://www.databricks.com/blog/introducing-omnigent-meta-harness-combine-control-and-share-your-agents 00:00 Meet Matei Zaharia: Creator of Apache Spark 00:34 The Origin Story: Escaping "Vibe Coding" 01:05 What is a Meta-Harness for AI Agents? 02:40 How to Hot-Swap AI Agents Mid-Session 05:06 Debugging 10x Faster with Parallel Agents 05:47 Why Big Context Windows Make Models Stupider 07:52 Inside the Architecture: Server vs Runner 09:07 Cloud Sandboxes: Stop Leaving Your Laptop Open 11:07 Beyond Static Rules: Contextual AI Security 14:12 Prevent a $1,000 Mistake: AI Spend Control 15:53 Why Databricks Open-Sourced Omnigent 17:59 Run AI Agents on Your Phone & Next Steps