How to run any agent in the cloud with Oz - Claude Code, Codex, or the Warp Agent

Added: 2026-05-21

[00:00:00]Hey, my name is Lili. I'm an engineer at Warp, and today I'm going to be showing how you can use Warp's cloud agents with other harnesses. Today, we're going to be implementing an improvement to the web app. So here you can see a bunch of Warp agent runs, and when I click into them, let's say I click into this one, I can see that it was run with the Claude Code harness, which is the new feature that we're demoing today.

[00:00:22]Unfortunately, we can't see at a glance out here which harness each one of these runs is using, and that's inconvenient. So what I'd like to do is add a new badge to the web app so that we can see up here which ones are using Warp's harness, which are using Cloud, which are using Codex, etc. All right, so we're going to start by creating a new run, and this is going to use the Warp server demo environment, which is using the dev web base image because we're doing some web development. I'm currently going to set it to the Codex harness for this, which you can select there, and I'm going to have it use my OpenAI API key, which I've already set up. I want this to run on Warp's infrastructure, and I'm going to add in a prompt here, and I'm also going to attach an image for the agent to use for some context. So here I have this example of what the state currently looks like, and this should help the agent get a better idea of the world that we're currently living in, so I'm going to attach that. Taking a look over this, it looks good, so I'm going to send this run off. All right, so now I have an agent session, and I can click into it, and in the app we'll see that Codex is running with my prompt, and it's gotten the attached image for the status badge example. So I'm going to let the agent keep running and come back in a bit when it has a PR. All right, so coming back to this run now, we can see that the run is done, and we have some artifacts. So it's reported back the PR that it's created to the Oz platform, and we can access it here via this panel. If I open this pull request, I'm able to review it, which is neat, but I'd like the agent to test this itself and upload some proof that it looks reasonable. So I'm actually going to go back to Codex and prompt it again.

[00:02:10]All right, so we have mocking infrastructure set up for our web app, and I'm going to ask Codex to verify its changes and upload some screenshot file artifacts so that I can actually see that what it's built has worked. And we're back. So we can see here, this is in the web app again, that this view is done, and we've uploaded these two new artifacts, these images that prove to us that the agent has implemented this correctly. So let's take a look. Opening this up, we can see that we have these Warp, Claude Code, Gemini, Codex badges, which is really awesome to see. And let's see the mobile version that it tested as well. Cool. This looks good on mobile too. If I go back to the session, I can see that it has used the web app to check these different cases and has created these images that it then reported back up to Oz, which is really awesome. And I have much more confidence now when I'm taking a look at this PR that Codex actually did a good job of implementing the change that I wanted.

[00:03:09]To take this demo a step further, I'm going to tag in a Claude Code agent to review this PR. And maybe a little overkill for this badge case, but provides an example of how you can really leverage multiple harnesses to get work done. So here on this PR, our Oz internal PR review integration has given me some feedback on the PR. And what I'm going to do is I'm going to go back to the web app and trigger a Claude Code agent to handle that feedback. All right. So here I am back in the web app.

[00:03:40]I've linked the PR and I'm going to ask the agent to address the feedback. This time I'm triggering it with Claude Code. So we should get a different harness's perspective on what we've done here.

[00:03:51]All right. So this session is now running. Let's open it up. Cool. So we can see that we have Claude Code running now with this prompt that we've passed in and it is pulling the review comments from the GitHub API. It's taking a look at what the Oz agent has said, and it is now making the changes to our PR.

[00:04:14]All right. So Claude Code is now done. It has fixed the issue and responded back on the original PR.

[00:04:21]And if I go back here, I can see I've responded, which is pretty cool. The last thing that I want to do just to close the loop fully here is I'm going to pull this branch locally and run it and make sure that the harness looks good. But I have a lot of faith after seeing Codex's screenshots and having Claude Code respond to the Warp harness's PR review here. All right. And here we are on my local server and we can see that all these badges look as expected. So this is how we use three different harnesses in the process of developing this feature. Definitely a little overkill for these badges, but hopefully a great representation of how you can use the Oz platform to leverage different harnesses at different points in the development cycle.