Meta-Harnesses for Secure, Collaborative Multi-Agent Coding with Omnigent on Databricks

Summary

Multi-Harness Unification: Consolidating diverse coding harnesses into a single interface simplifies session management and eliminates context-switching costs across different AI tools.
OS-Level Sandboxing: Implement OS-level sandboxing policies that prevent autonomous agents from making unauthorized system or directory modifications and accessing sensitive data.
Cross-Model Orchestration: Orchestrating workflows across different AI providers and model families produces higher quality code reviews and automated issue resolution than single-model approaches.

Meta-Harnesses for Secure, Collaborative Multi-Agent Coding with Omnigent on Databricks

Talk by Brian, Engineer, Omnigent team, Databricks
A meta-harness is a layer above coding harnesses like Claude Code, Codex, Cursor, Pi and OpenCode that gives you one interface, shared work and enforced security. Omnigent, built by Databricks, addresses harness sprawl, terminal-bound workflows and risky agents with a unified interface, mobile control, sandboxing and custom policies.
An Omnigent engineer explains how Databricks went from no coding harnesses to many, the switching cost between them, and cited research showing Claude reviewing Codex work beats Codex reviewing itself. The video covers the architecture: a server that stores history and policies, hostable on Kubernetes, Railway, Vercel or Databricks Apps, plus a runner that wraps the bare Claude or Codex command. It demonstrates contextual policies that deny or ask for approval, such as blocking an SSN, OS-level sandboxing with bubblewrap on Linux and Seatbelt on Mac, cloud sandboxes on Modal, Daytona, EC2 and Kubernetes, and repro and resolve agents that merged dozens of PRs.

Chapters

FAQs

Why do developers face operational friction when using multiple coding harnesses?

Switching between different harness UIs and model ecosystems creates fragmented workflows, context loss, and security risks. This video presents a unified orchestration layer solution that consolidates tools.

How does OS-level sandboxing prevent unintended autonomous agent actions?

OS sandboxes using technologies like Bubblewrap or Seatbelt enforce customizable policies restricting agents from unauthorized modifications. This video details practical sandboxing approaches for securing agent execution.

What are the benefits of enabling real-time collaborative AI coding sessions?

Team members can co-drive AI sessions, streamline on-call handoffs, and accelerate pull request generation with human sign-offs. This video demonstrates how collaboration improves agent workflows and team efficiency.

Can different AI models be combined in a single agent workflow?

Yes—orchestrating cross-model workflows that combine output from different providers produces superior code quality. This video shows research proving mixed-model approaches significantly outperform single-model systems.

Full transcript

`00:00` Introduction

Hello everyone. My name is Brian. I'm an engineer that works on the Omnient team at Data Bricks. Today I'll be talking a little bit about what Omnient is, how it can maybe make your agent coding a bit more secure, and how all of you could benefit from using Omniant. Uh so I'm going to start a bit with some of the problems that we've been encountering at Data Bricks, uh and as an industry as a whole. So I'm sure you guys know there are many many harnesses out there today. uh cloud code, codeex, cursor, pi, open code, kimmy code, quen code, as many codes as you want.

`00:15` Why are there so many coding harnesses?

Uh and at data bicks, we've actually rolled out quite a few of these. Um it's hard to believe that just one and a half years ago, there was no coding harness available at data bicks. Uh people were experimenting with copy and pasting into chat GBT, even though it was against company policy at the time. Uh but then you know we finally rolled out cursor tab autocomplete uh goose CLI in the early stages to help us automate some unit testing uh and investigations.
But I think things really took a turn in probably June of 2025 at data bicks where folks weren't just using it to assist your coding. It was actually heavily assisting or maybe even writing a lot of the code that you ended up merging. Uh and I'm sure all of you have only experienced more of this since then. One particular challenge we had at data bicks is how do you know which harness you should be using? When cursor first came out, it was really really cool.

`01:19` Which coding harness should you use? Speed, quality and cost

The tab autocomplete was way better than what I could be typing in terms of how fast it was. Uh but then when cloud code came out, it was even better. Uh I could talk in natural language. It would produce a bunch of code for me. yada yada. I'm sure you guys know this phenomenon. Um I think today there's a bit more par between cloud code, codeex, and cursor. It's not really clear which option is the best for what I'm doing.
There's trade-offs to make between speed, quality, and cost uh that are all kind of hard to weigh. There's been some research that shows that when you mix and match, you get the best quality. Uh I know I just Googled for some paper that showed something like this where if you use claw to review Codex's work, it's much better than Codex reviewing Codex's own work by I think in this case uh 18% if I can do my math right.
Uh, but there's a bit of a switching cost here between going from the claw tui to the codeex desktop app or maybe the cursor SDK. There's a bunch of different surfaces and ways to interact with each of these harnesses that isn't necessarily the most straightforward. Those skills and MCP servers and memories that you develop with one harness are not automatically transferred across when you switch just because your coworker told you it's really cool.

`02:35` Moving beyond the terminal: organizing agent work

Another problem we're trying to solve with Omnigent is to get you outside of the terminal. When Cloud Code first came out, I was stuck inside the terminal with many T-Mux PES. I'm sure you guys had similar experiences. You know, one tab was five T-Mux PES all doing something slightly different. I had a bunch of forks uh and summarized compacted conversations. And every morning when I came back to my computer, it was an exercise in figuring out which session belonged to what task I was working on.
Um, and inside of data bricks, we've actually seen people use a bunch of different tools. On the left here is open UI. It's like a open source project that we forked and developed internally. Uh there's also cursor and the codeex desktop app both great agent orchestrators as well. And all of them are kind of solving the same session organization problem where I have all these t-ox pains but in reality with the representing our units of work.
How do I organize and show these to the user in a way that makes sense and is logical? Uh and then one final problem. We got dangerous agents. I'm including an example that's maybe a bit more relatable to us here where you know sometimes I let my cloud code run on yolo mode or my codecs run on yolo mode and you know this guy right after soul released I think he posted this it got quite a few views uh 6 million and the codeex creators responded and offered a bunch of different stuff to him uh you know as compensation basically for soul going off the rails but if I wanted to try soul and I'm used to running cloud code in yolo mode maybe I wouldn't have known that this is a potential consequence.

`03:37` Risks of dangerous agents in yolo mode

These LMS are so smart. I would have never thought it would delete everything on my Mac. Uh so we've got those four problems and we're trying to solve all of them with Omnigent. As a developer, I feel like whenever I hear about a new dev tool, I'm always asking myself, what does it get me? Uh what does it get my particular workflow? How does it make my dev experience better? Um and hopefully between these six value props, something will resonate with you.

`04:30` Six benefits of a unified meta-harness

I think the biggest is the unified interface for all the coding harnesses. So this is between cloud code, codeex, cursor, pi, open code, kimode, you name it and we probably have an integration for it. Uh and if we do not, we also support community harnesses. Uh I know Atlassian added one with rovo. We're about to add another one for Klein. Uh so super exciting time for the open source community there. There's also full mobile control of every session. uh at least for me whenever you know I'm walking home from work I've kicked off some work on my dev box I can check on it as I walked home and you know when we rolled this out at data bricks I feel like people talked a lot about work life balance and this is either going to affect your work or your life depending on which one you're looking to balance um and you know a third benefit of omnient here is powerful sandboxing and custom policies this is how you're going to make your coding agent more secure and not do things that you don't really expect.
You're also going to be able to create agents that are composable across model families. So this means maybe having a clawed orchestrator that talks to a PI sub aent or a codec sub aent uh and vice versa. You can achieve those same results that they found in the research paper with higher quality code by running it through different model families super easily. You can also share your sessions with teammates. Uh, I frequently use this to work on stuff like on call where while I'm on call I probably know like 80% of all the issues that are going on and that last 20% I need that subject matter expert or the person who actually created the feature to come in and help me finish the fix.
I just hand it off to them. They get to work and I'm we're able to ship a PR as if we're pair programming right next to each other. RTL works down in Mountain View, so sometimes we don't get to see him too often, but as soon as we hop on a video call and we hop on the same session, it feels like we're sitting right next to each other. Uh, and we also get you cloud agents. I know they've been all the hype recently. Uh, you know, Devon, Cursor, Codeex, they're allowing you to launch cloud agents.
Uh, but it's all on their infra and you're not able to customize too much of those compute environments. It might not be inside of your VPC. Uh, and Omni is going to allow you to do all of that stuff open source. So, how do things actually work? Uh, taking a step back, I think Casey also showed this in her demo, but we have a server. And this server is the thing that will actually store all of your session history. It's going to allow different clients, including the people that you shared the session with, to access your agent sessions.

`06:51` How Omnigent works: server and runner

Uh, it has policies, MCPs, skills, artifacts, etc. Um, this can either be your own K8 cluster that you're hosting. It can be hosted on Railway, Verscell, Data Bricks apps. Uh, and then we also have a data bricks manage server if you guys are interested in that. Um, and then there's also a runner piece. This is the guy that's going to actually run your agent. So all of those cloud code skills and memories that you've accumulated over your last year and a half of vibe coding, uh, all of that's going to be accessible to Omnigent as well.
We just run the bare claude or codeex command under the hood. Um, and so there should be minimal setup required to get going. The mobile Omnient app allows you to drive stuff from anywhere and it's available today in both the iOS app store and the Android Google Play Store. We have Omnigent contextual policies. This is that PII policy where it told me that it can't send my SSN because it's, you know, private. Um, and these policies are a bit more complex than standard hooks.

`07:55` Contextual policies: deny or ask for approval

There's the ability to ask for approval before straight out denying it. So in this case, the SSN one was straight up denied. However, if I had sent maybe my credit card information to an agent because I explicitly wanted to be buying something, it will ask me for permission before denying the tool call. Um, and these are policies that are probably pretty specific to me. uh as I use these agents for my taxes or something else that's personal, I don't necessarily want 100% of the things that apply for the company to apply to me as well.

`08:32` OS-level sandboxing and cloud sandboxes

We also have OS level sandboxing. So this is what you saw where I'm trying to get outside of that working directory into my home directory and I'm getting denied deterministically. I'm not just begging the agent not to be, you know, elusive or, you know, wander outside. Uh it physically cannot access those files. On Linux, we use bubble wrap. So that was my dev box. That's why we saw BRP in the uh tool call error. And then we also have something called seat belt if you're running this on Mac.
Alternatively, you can also use cloud sandboxes, which is those cloud agents that we were talking about. We support integrations with modal, Daytona, uh Amazon EC2, Kubernetes. Um and these will allow you to run on physically isolated incidents. uh sorry physically isolated compute that will prevent uh the agents from having access to anything sensitive and this sandboxing works across every harness. Uh as we were rolling stuff out at data bricks we originally came up with something for claude and then we came up with something separate for codeex and then we came up with something separate for cursor.
Uh with omnient hopefully the approach is all you need to do is swap out this one line uh between cloud SDK and codeex uh just one line will get you the same exact sandboxing protections uh that will make you feel safe with your agents. We also have custom agents. So this is what I was talking about earlier where you can orchestrate across different models or different harnesses. Um, this allows you to easily send over maybe a workflow that you've been working with to your teammates so that they can also follow your review agents.

`10:02` Custom agents: repro and resolve agents

It uses Claude and Codeex, combines the results, and then gives you a consolidated review. For our team, we actually have a bunch of custom agents called the repro agent and the resolve agent. These guys comb through all of our GitHub issues. We make sure that there's a detailed reproduction instructions for every single issue that's filed. It gets triaged as a P 0, P1 or P2. And then we separately also have resolve agents that run 24/7 that look through all the P 0 and P1's tries to fix them if there's a standard uh reproduction and then includes screenshots and video recordings of the fixes on the PR and then requests a human for review.
We're not quite at the point yet where the agent merges by themselves, but I think we're getting close. Uh we just turned this on last week and there's already been 38 PRs that have been merged and you can see here there's 67. and they're open on the repo right now. Uh I think we closed 100 something issues over the weekend, mostly just by people clicking stamp after the resolve bot puts up a PR. This is the collaboration that we were talking about.
Uh I can allow other people to comment on my sessions. I can comment on their sessions. Um and we can co-drive agent sessions together. Some internal screenshots. Uh mate, our CEO was using Paulie just about 20 minutes to figure out how we can get the desktop app going. You can see now I use it every single day. So, you know, Paulie did a great job there. Uh, Aarva is one of our early users at Datab Bricks. Um, the screenshot might be a little old.

`11:16` Co-driving agent work with teammates

Um, and he's been using it as a replacement for the other tools that we have at data bicks because of the ability to work on your phone and the terminal view and the sharing features that we have. Uh, Peter from our docs team works super hard on the docs, but at the end of the day, he still needs a bit of engineering review. uh he invited us to an Omnient session where we're able to talk to one another directly and edit the docs as if he's the one talking to the agents.
We're getting this deployed inside of other companies as well. Um I don't think we're allowed to talk exactly about which companies they are, but there's a few digital native businesses that are rolling this out inside of their own engineering orgs. Uh and if you want to use this for personal use, it allows you to use a bunch of different of your cloud subscriptions. Uh I know the tokens are subsidized. Um, and multiple $20 subscriptions are more powerful than one $100 one.
Uh, but yeah, you can fork across and then just run with different subscriptions over and over. The community is excited. We have just over 9,500 stars today uh and 1500 forks. There's a bunch of PRs over time. We're working on making sure that they get merged in a timely manner still. Uh, but it's a really growing community and we'd be really excited for all of you to join us. Thank you guys so much. Here are some links where you can reach us. a QR code for you to join our Discord.

`12:31` Community and getting started

Um, you know, people from the digital native businesses that we're talking to uh talk a lot in there and we're also pretty responsive.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.