Full transcript of Introducing Omnigent: an Open Meta-Harness.
Source: https://www.youtube.com/watch?v=3Xv0VjYhpCk
Welcome to the stage, Databricks co-founder, Matei Zaharia.
Hi. All right. Hi everyone. Welcome to the second day of keynotes. So, I’m excited to kick off the morning today and talk about a new open source project we launched over the weekend. We just released it Saturday. This is new layer from our engineering team for working and building with agents and it’s based on patterns and challenges that we saw coming up everywhere where people are using agents and we realized that we’re going to need this new layer on top of the current agentic stack to make it really effective. So, I’ll talk about what that means, but we call it a meta harness - harness of harnesses, if you know what an agent harness is.
So, as everyone probably knows, agents are really starting to work. Coding agents are the first type that were really successful. Pretty much every developer has probably used them at this point, but the same is true in other fields and you saw with Genie and with other agents that you can make very effective agents. And especially because coding is a lot easier now, you might wonder do we even need agent infrastructure at all? If I don’t like the current agents, can I just write code my own? So, it seems like maybe we’re kind of done with at least simple software and we’re set. But actually when you use agents, it can be quite annoying if you’re a manager, it can be very scary because of security issues and stuff like that. So, the world isn’t really perfect yet.
What’s an agent harness? The agent harness is basically the software around the language model that is the agent’s interface to the world. So Claude Code as a coding agent - there’s Claude the model that just outputs tokens, and then there’s Claude Code the harness, which connects it to files and a user interface and security permissions. And then there’s Codex, and any custom agent you make - you have some kind of software around it called the harness. And because it’s so easy to build harnesses, everyone is building them. There are so many agent harnesses - just the popular ones for coding, and then custom agents with agent SDKs. It’s a ton of these out there.
The problem? Agents sometimes have to interoperate, and there are common challenges that the harnesses have where it helps to have a common layer that lets us interface across them. If everyone is just vibe coding their own thing, they will still be incoherent.
Three main challenges:
First, composing agents. When engineers at Databricks do coding, they use many coding agents at the same time. If one is stuck, they ask another. They give them different inputs, work on different things in parallel. How do you compose agents with different harnesses? It’s easy within one (Claude Code or Codex), but what if you want to use both? There’s many patterns emerging - loops (agents supervising each other), advisors (small cheap model with an expensive model directing it), and other composition patterns.
Second, collaborating across humans and agents. If I’m using one coding agent and someone else is using another, how do we collaborate? Each agent has its own harness, its own app, and it’s stuck there. Why can’t I hand my session to another engineer and have them continue with the full context?
Third, control. Agents can be hacked through prompt injection. They can make mistakes like deleting databases. They can cost a lot. Current permission prompts are useless - “Do you want to run this hundred line Python script? Yes or no?” Just hoping users will review every action isn’t enough.
We think we need a new layer above the harness level to manage and build with agents. It should be open so anyone can plug in their own harnesses, workflows, and tools. We call this layer the meta harness. That’s what we just released with Omnigent.
How does Omnigent work? At the bottom you’ve got existing agents - it works with coding agents (Claude Code, Codex, plus anti-gravity and coarser added since Saturday). It also works if you’re building custom agents with something like OpenAI Agents SDK. The interface is simple - you can send messages in, get a stream of events out, call tools.
Omnigent has two pieces. First, a runner which wraps around your agent and enforces security and controls by sandboxing it. Second, you can optionally connect Omnigent to a server where you get collaboration and central policy features - central history, agent sharing, MCPs and skills centrally, security policies, artifact sharing. On top of this, any agent gets a uniform interface - web UI, native app, mobile, and REST API. Omnigent is open source and does not require Databricks in any way.
What does this give you?
For composition, you can combine agents that use multiple harnesses and LLMs. Write a YAML file to define a multi-agent workflow. Switch agent harnesses while running.
For collaboration, we support live collaboration. You can invite other people to read and write to your session, talk to agents, and comment on files in real time.
For control, we came up with a much more powerful model than what you have today. Today most agent tools have static permissions - for each tool call, allow/deny/ask. We support contextual policies where permissions vary based on what the agent did before. If the agent had a confidential document, maybe we don’t let it send email outside. But if it only had public marketing docs, it’s okay. We also have policies on cost - tell it not to spend more than $5. We built a flexible OS sandbox to protect the agent from harming your machine.
Demo: Started Claude through Omnigent on a Databricks dev box. Same standard Claude interface. But you get a web UI too - view files, see diffs, leave comments, tell the agent to address all comments. There’s also a native app with notifications. You can switch agents mid-session - fork to Codex with the same history. Or run Polly, a multi-agent supervisor that uses both Claude Code and Codex for subtasks in parallel.
For collaboration, I invited a Databricks engineer to my session. He could see everything, leave comments, and send messages to the agent. I literally called him and he reviewed my readme changes live.
For policies, we track cost per agent and allow contextual budgets. Set “ask me if spending more than 20 cents” and the agent pauses for approval when triggered. Eventually hits a hard stop at the configured limit.
Why open source? Working with agents has to go above the level of a single harness and SDK. This meta harness layer benefits from being open the same way data formats and sharing protocols do - it’s an ecosystem where people have to agree on a way to do things. We use it a lot internally. Since Saturday we already got great community contributions.
There’s a lot more coming - more ecosystem support, auto optimization of agents. We really think AI needs a layer above the harness. It makes it way easier to work. Our team (engineers, product managers, designers) are all using it. Check it out.