Omnigent: Databricks’ Meta-Harness for Unified AI Agent Management

Generated: 2026-06-16 · API: Gemini 2.5 Flash · Modes: Summary

Omnigent: Databricks’ Meta-Harness for Unified AI Agent Management

Clip title: The Meta-Harness: Why Every AI Developer Needs This Author / channel: Prompt Engineering URL: https://www.youtube.com/watch?v=141biWM1mlE

Summary

The video introduces Omnigent, an open-source “meta-harness” for AI agents developed by Databricks, designed to address the current fragmentation and inefficiency in working with multiple AI models. The presenter highlights that users often juggle several AI agents (e.g., for coding, review, chat, search, documentation), each with its own capabilities, memory, and user interface, operating in isolation. This necessitates constant manual copying and pasting, leading to inefficiencies and “vendor lock-in” where developers are forced to re-engineer their systems when better models or SDKs emerge. Omnigent proposes a unifying layer that standardizes the interface across these diverse agents, allowing them to work together seamlessly.

At its core, Omnigent consists of three main pieces: agents, a runner, and a server. Agents can be proprietary (like Claude Code, Codex, Gemini, Pi) or custom-defined via YAML files. The “runner” component encapsulates any agent into a uniform, sandboxed session, ensuring reliability and security. This session is then exposed through the “server,” which adds crucial functionalities such as persistent history, a catalog, policy enforcement, artifact management, and skills. The server is PostgreSQL-backed and deployable across various environments (Docker, cloud sandboxes), making the unified session accessible from any interface—be it a terminal, web UI, native app, or mobile device—allowing users to start a task on one device and continue it on another without losing context.

Omnigent unlocks three key capabilities. Firstly, Composition, allowing users to swap or combine different AI agent harnesses without rewriting code. Agents are defined simply via YAML, making it a one-line change to switch between underlying models. It also enables multi-agent workflows, exemplified by “Polly,” a built-in orchestrator that plans tasks, delegates coding to different agents, and facilitates cross-vendor code review (e.g., Claude Code writes, Codex reviews) to mitigate biases. Another agent, “Debby,” serves as a brainstorming partner by soliciting responses from multiple LLMs (like Claude and GPT) and even prompting them to debate each other for refined answers.

Secondly, Control is built directly into the Omnigent layer, not merely suggested via prompts. Every action an agent attempts (like installing a package or pushing code) passes through an enforced gate (allow, deny, or ask for user approval). This policy enforcement is stateful, meaning rules can depend on the session’s history and apply to critical aspects like cost budgets, risk scores, data access scopes (repo/file), and PII scanning. Furthermore, agents operate within an OS sandbox, restricted to only accessing explicitly allowed files and networks. Importantly, the agents never directly handle sensitive API keys; Omnigent injects these secrets only through approved egress proxies, significantly enhancing security even in “YOLO mode.”

Finally, Collaboration is a core feature. Live sessions can be shared via a link, allowing teammates to watch the agent’s progress in real-time, chat with it, or even “co-drive” by having their messages run on the host’s machine. Users can also fork conversations, enabling independent exploration from any point. Omnigent aims to transform AI agents from solo tools into collaborative team members, providing a robust, controlled, and flexible framework for developing and utilizing AI solutions. While still in its alpha stage, its open-source nature promises rapid evolution and community contributions.

Video Description & Links

Description

Thanks to Databricks for early access and making this video possible through their sponsorship.

You use Claude Code, Codex, and Pi separately — copying context between them because no agent can see the others. In this video I break down the meta-harness: a single layer that sits above every harness so all your agents share one session, one history, and one set of policies — and even review each other’s code across vendors (Claude writes, Codex reviews). I walk through OmniGent, the Apache-2.0 meta-harness Databricks just open-sourced, and build a real app with it live.

LINKS: Blogpost: https://omnigent.ai/ Github: https://github.com/omnigent-ai/omnigent

My voice to text App: whryte.com Website: https://engineerprompt.ai/ RAG Beyond Basics Course: https://prompt-s-site.thinkific.com/courses/rag Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0

Let’s Connect: 🦾 Discord: https://discord.com/invite/t4eYQRUcXB ☕ Buy me a Coffee: https://ko-fi.com/promptengineering |🔴 Patreon: https://www.patreon.com/PromptEngineering 💼Consulting: https://calendly.com/engineerprompt/consulting-call 📧 Business Contact: engineerprompt@gmail.com Become Member: http://tinyurl.com/y5h28s6h

💻 Pre-configured localGPT VM: https://bit.ly/localGPT (use Code: PromptEngineering for 50% off).

Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0

00:00 - The Problem: The “Agent Box” Trap 01:40 - What is a Meta-Harness? 02:22 - The Architecture: Runner, Server, and Sandbox 03:50 - Multi-Agent Composition 04:15 - Meet ‘Poly’: Cross-Vendor Code Reviews 05:03 - Meet ‘Debbie’: The AI Debate Partner 05:45 - Governance & Safety: No more “YOLO” runs 07:08 - Real-Time Collaboration Features 07:32 - Step-by-Step Setup & Installation 09:12 - Live Demo: Building a Gemini Web App 11:40 - Monitoring Costs & Token Usage 13:10 - Final Results & The Future of Orchestration

NemoClaw Knowledge Wiki

Explorer

Omnigent: Databricks' Meta-Harness for Unified AI Agent Management

Omnigent: Databricks’ Meta-Harness for Unified AI Agent Management