The Agentic Harness: What Makes AI Agents Work in Production | Snowflake Dev Day 2026

[00:00:05]Quick show of hands. How many of you have asked AI to write code for you this year?

[00:00:12]Everyone should be raising your hands right now. Keep your hand up. Come on, keep your hand up if that code actually ran on the first try.

[00:00:22]Fewer. Okay. Our next speaker is here to talk about why that gap exists and what we can do to close it. So please join me in welcoming Roman Huitt, head of developer experience at OpenAI.

[00:00:38]Thank you.

[00:00:41]Good afternoon everyone.

[00:00:44]Super excited to be here. I'm I'm excited to talk to you about the rise of the Aentic coding and the Aentic harness today. And if we get started to look back a couple years, it's going to be no surprise to any of you that the way we build software has completely changed, right? We started with a phase of code complete and we were quite happy with having an IDE that would kind of autocomplete the lines or even some functions. But then quite quickly we went to pair programming and this was this era where like the IDE could actually do more for us. it could write a feature end to end and that alone was pretty remarkable.

[00:01:24]But now we've entered the third phase which which we call agentic delegation.

[00:01:28]It's the idea that like we're not even writing code anymore but rather delegating to agents the task to build this software.

[00:01:36]And so the way coding is accelerating has also changed the way we think about building software, right? Because for decades we thought about coding and the ability to write the software as the limiting factor. You know the engineers have always been the scarce resource and we had to spend a ton of time on like writing the code but also carefully reviewing it. But the thing is like as we planned on this now agentic coding has changed the whole pipeline right because that assumption has changed it completely changes the way uh you can build software because you can write so much more code and so now the constraint has moved around the writing it's like how can I actually do everything before and after how can I do more planning more reviewing more security checks and so on and so forth and so when we're building uh codeex at openai we're thinking Think about the entire software development life cycle.

[00:02:31]We want to make sure it's not just this thin pipe that's very narrow but actually increasing the pipe so you can end to end accelerate the entire journey. Like in building in that phase the engineers delegate the implementation to to the agents. In review we have codeex that helps you inspect the changes and have a human review as well. And in deploy and operations, we're also thinking about how cloud agents and automations can really ensure that like you can investigate issues at scale and and keep moving uh through the system. And the goal is not to remove the humans at all in this system. It's more to empower them in this new phase of agentic delegation across all of these different things that they have to do. And if Codex is going to operate across this entire software um uh development life cycle, we actually need more than a great model.

[00:03:24]And so first we realized what that we needed a better product surface, a way to really have a command center to delegate work to agents. You know, the terminals were never meant for coding.

[00:03:35]At best, they were built to run a few commands, but like writing code in the terminal does not make much sense. you really want to do a lot more than that and monitor all of these agents that you're spawning. And so that's why we built the Codex app, a new command center to delegate real work to agents.

[00:03:51]Actually, quick show of hands maybe who has used the Codex app so far.

[00:03:56]Quite a few of you, that's awesome.

[00:03:57]Well, for those of you who have not tried it yet, I'll I'll show you some live demos uh towards the end. And with this new surface, we are actually seeing a lot of great momentum. We now have more than five million weekly active users on the Codex app. Many great teams using it. And it's really moved beyond like just a simple app to to write some code. It's really something where a lot of developers but also everyday workers are using to delegate real work to uh to these agents.

[00:04:27]And so for us the question is how does that work and where do we start? Well, first we always have to start with the model as well. And so GPT 5.5 that we launched now just a months ago is our frontier model. And what we care about is not just that the model is able to do complex reasoning, complex tool calling and high quality outputs. What we care about is that it it should also be very token efficient. One thing we say internally for instance is that GPT 5.5 only hates one thing and the one thing that it hates is to waste tokens. We really want to make sure that we jam-pack the most intelligence possible per uh token units.

[00:05:07]And that's very important, right?

[00:05:09]Because it's not just a cost metric.

[00:05:10]It's also changing how you think about the kind of useful work your agents are doing. If your agent, let's say, is inspecting the codebase or maybe it's like running commands or it's like, you know, uh revising its own plan, you want to make sure the agent is really useful and spending tokens the right way.

[00:05:28]There was also this new benchmark that came out just a couple weeks ago I believe called deep suite because I think a lot of the SU benchmarks have been quite saturated in the industry now because it's really hard to have models that have not already seen in their training some of the tasks and what's interesting about this one is that they try to make the prompts much shorter but the outputs of the model much more complex to to uh to run and on this one GPT 5.5 is not only the best at quality but what we love to CL. So it's like it it does so with like a third of the tok to tokens often. So that's that's pretty cool for uh for um builders using it.

[00:06:06]But between a new surface and this new model, you also have a very critical piece which is kind of the topic I wanted to bring up today and that's the harness. You know, you have the model layer which does the fontier reasoning.

[00:06:18]You have the product surface to think about like you know your cloud agents and your enterprise controls but in between you need to have this harness layer. And what we mean by that is this idea that like you really want to have all of the context, all of the tools, the way for the model to verify its own work and verify this like this work with multiple other tools and also an execution environment that is safe and secure. So one question we get often for instance is like how um did you guys like u you know spend a few more weeks to ship the app for Windows before Mac OS and part of the reason was actually to work on this sandboxing and this execution environment for for the model and of course on top of that you also want to make sure that the output is re reviewable by a human and that's collectively the environment that we kind of you know refer to as the harness.

[00:07:09]A few things I wanted to call out about uh Codex and especially the Codex harness which I think is is quite unique. The Codex harness is completely open source. It's part of our you know OpenAI/Codex uh GitHub repo. The app server and the CLI is also like open source. So developers like all of you can inspect what's in there and even extend it and adapt it to your needs.

[00:07:29]You can also use the codeex harness anywhere. So you had Peter from OpenClaw today and like now you can use your your chat GPT subscription in Open Claw and OpenClaw now supports the Codex harness as well when you're kind of building your agents.

[00:07:43]We also ship fast. I think I actually have an update once again to do on my Codex app over here because the the team is cooking a lot. Another thing that's very interesting about Codex is that safety as I mentioned is part of the architecture. And so on the right side, I won't bore you with the details, but this is all the work we had to get right in order for the cloud the the the coding agents to perform correctly on Windows. And uh we want to make sure that we have all of the sandboxing done right so that whether you're on Mac OS or any other OS like Windows, the uh agent can only access the things that you uh that you approve.

[00:08:18]And on top of that, we also try to make sure that like whatever is happening in the Codex harness, the tool it's calling and the progress on any task can also be playful. That's why we launched the codeex pets for instance. You know, having these long horizon tasks that run on your desktop, but you can monitor them really at a glance. And lastly, we're trying to push also the kind of tools this harness can use. So, we launched a research preview of something called Chronicle, for instance. And chronicle is this idea that like Codex can actually observe your desktop as you're doing and conducting your work.

[00:08:48]And based on that, you can ask Codex actually what have you noticed on my work? Is there anything that you could take off of my plate by watching me do it? And so all of these things are part of the tools that we bring into the harness.

[00:09:01]And along the way, what we've learned that's very interesting is that Codex is now much more than just for coding.

[00:09:07]Because it turns out when you have such a great harness that's able to call the tools have all of the uh execution environments and as part of these tools you add let's say an inapp browser for testing for verification you add things like memories what that enables is that any kind of agent for any kind of work can actually be done using the same harness and that's why for instance at openai virtually everyone is using the codeex app including our finance team our CFO is even using the codeex app even though she's obviously not coding all day.

[00:09:40]And the other thing we've done and um we've announced actually this week is that the Codex harness and the Codex features are making their way into Chad GPT as well. We want to make sure that anyone using chat GPT can actually unlock this power of like agentic coding but also agenting tasks and this world of delegation. In fact, what people have maybe not noticed yet is that up until GPT 5.3 codeex, we had a very specific coding line for our models. But as of GPT4, GPT 5.4, sorry, now 55, we have re reunified all of this so that it's like one mainline model for all kinds of work coding but also non-coding tasks. We've already launched some new features in CHP on top of the codeex harness. So the final uh kind of unification story is to kind of bring all of these surfaces together so it's even easier to reason about and at the same time what I think is very interesting is what's happening for platform companies you know from Twilio to stripe I spent many years at stripe before openai what does that mean for them well what's very interesting to me is that when agents starting to write and build these integrations snowflake as well uh your product has now as now a new user and that new user is kind of an agent conducting the work. So just as an example of a few things that you have to think about now is like you know your docs are obviously very important but they're mostly read by agents conducting that work. I used to spend you know many days or weeks sometimes at Stripe building sample applications. Well those sample apps are maybe not that critical anymore. What you want instead is to have some great skills, some great plugins and sure enough the model will be able to build some of those on the fly. So there's quite a few things changing including for API keys letting an agent actually sign up for a platform not just a human. And so that's kind of what we refer to as like agent experience in the industry. I was actually at Stripe when we launched the Stripe CLI and they published this chart recently which I find amazing. Uh the Stripe CLI has been now around for seven years, but you can see the rise now because most of the users of the Stripe CLI are actually agents like Codex writing the code and interfacing with the Stripe API.

[00:11:53]And of course, if you're building enterprise agents, I think the context is also very important because it has to be governed and so it needs to have access to the right information. Also needs the shared definition, the right permissions, the right lineage, the right audit traces and all of that. And that's why we have the Snowflake plugin as part of the 100 plus plugins we have in Codex. So if you're building with Snowflake, you can just simply add the Codex plugin to your to your app and start uh benefiting from it. And so what that means is that with something like the Codex app, we're really seeing something where what used to be conversations or threads that's now really becoming goals. It's this idea like you can raise your ambition. You can really delegate very complex tasks to these agents whether they're coding or non-coding tasks. and just delegate that work and just monitor it from a surface like the Codex app.

[00:12:43]So with that, maybe I'll spend a few minutes to give you a quick tour of the Codex app and also what's interesting about like the harness and how it can manifest. So I'm going to try to switch to my screen. Sounds great. Um, so I'm going to kick us off with a very simple task, but instead of typing, I'm just talking to Codex these days. So I'll say, I'm on stage right now. Can you uh figure out the details from my calendar and give me kind of the overview of my day?

[00:13:08]Boom. So with that, it's a very vague prompt, very simple one. But what's going to happen is that unlike a typical chatbot, now Codex has to figure out like what is the context here. So right away it's pulling details from my Google calendar. It's searching from events.

[00:13:23]It's found five of them already and it's going to try to figure out what uh my day looks like, right? And that's very interesting because now boom in 15 seconds I have everything that's going on about my day by pulling details from from Codeex, right? Very cool. But of course you also want to have Codex that's like proactive for you. It can perform tasks for you behind the scenes.

[00:13:45]So here for instance if I switch to like the chief of staff, this is a very simple prompt that I sent like hey every morning at 7 a.m. I want you to pull everything from all the sources that matter to me and I want to have a brief of my day. And sure enough, that's all it takes to create an automation and then you have like every morning something that's like relevant to you for your day. So you really just talk to Codex and thanks to the harness connecting to all of these tools, you can see on the right hand side the kind of tools it's been using to actually produce this output.

[00:14:15]Now I wanted to show you a few things that I find very interesting when you start using uh the Codex app to conduct kind of any kind of work. Imagine I'm like meeting with customers and imagine they want to create like you know new kinds of of decks or new kind of like brand campaigns. We have an amazing image generation model. And so what that means here is like for instance creating a deck sure enough no problem to pull a deck from box uh you know pull details from like a template pull all of the context from the Salesforce on the slack and boom on the fly I can actually create a new PowerPoint. So very interesting when uh you go meet customers. But let's take a look at this one for instance. Let's pretend I work for a bank for a second. And let's pretend I'm trying to build a new mobile app in Swift from scratch. Well, of course, I designed my app uh using uh using codeex um right here. So, you see a few screenshots of it. But you can also create like all sorts of things including like what the website could look like. I like this one. This one is pretty cool as well. And of course, you get GPT 5.5 to implement these things.

[00:15:16]You can even do complete mood boards like you know. So just to show you like the the breadth of like what image generation can actually create uh when you're starting to build and if you're a front end or full stack engineer I genen is actually an amazing way to start designing a product and then turning it into uh reality.

[00:15:34]But now the one cool thing I wanted to show you is like let's assume we've launched uh this mobile app and let's assume that we've connected the snowflake uh new plugin that we built with the snowflake team. I can just simply say send this very trivial prompt which is hey take a look at how the mobile app is is doing on snowflake find the right context and the right data from it the launch and make me a dashboard per country well so this is coming straight out of the model uh using a new plugin that we built called data analytics but I don't even have to mention the plug-in at all I can just let Codex like understand it's a data question and then based on that I can take a look at a few things but not only do I have like the onboarding funnel and all of these metrics at a glance. I even have like the raw intelligence right there from Codex telling me that something looks off in Mexico on the chart and it sounds like Codex even has a guess from the snowflake data as to like it might be like due to a phone format in the kind of onboarding flow and it's true. It's pretty pretty down here. I'm going to show you one of my coolest one of my favorite features of Codex which I think is the coolest uh that we launched in the past couple weeks and it's called appshot. So imagine I'm bringing the app now uh front and center over here. I can just press command command on my Mac and I can just say can you review the onboarding flow in Mexico and then when you're done you can do it for France.

[00:16:59]So an appshot is actually a better smarter screenshot. So, it's basically taking a look at the app in front of my screen, but instead of just taking a screenshot, it also understands the context and it's able to drive the computer for me. So, here you'll be able to see that like the model is actually clicking. It has its own cursor which is not mine. My cursor is still here and I can still go work on my computer, but Codeex is able to actually go through the onboarding flow and click around filling out forms. And to my knowledge, this is the only implementation in the industry of what we call background computer use. So I can keep on, you know, showing something else on Codex.

[00:17:36]And meanwhile, I'm going to let the model figure out like how to click and how to reason through the different things. And once again, this is very much bound to the codeex harness being able to have the right environment, the right tools, and in this case, it's able to drive the iPhone simulator itself and understand what's happening on the screen. And sure enough, it found the bug. It sounds like if you're in Mexico, even though you have a valid number, you seem to be stuck. So that's an amazing way to think through like, wow, like what are the possibilities of such a harness when you have all of these tools at your disposal to build? Not only do I have all of the plugins to connect to my context in the cloud, but if something requires my computer to be driven by the model, the model can also ask me for my permission and start driving. So I can let it run in the background and then sure enough for other countries it should work fine. It sounds like it was very much isolated to Mexico. So there's so many features. I won't have the time to show you all of them. Maybe uh a couple uh that I like. One is the the new goal command. So imagine we have now launched this app and imagine we want to be very ambitious in terms of like crushing our backlog to make a new app.

[00:18:46]I can just say, can you take all of the all of the backlog from the launch yesterday and start a new PR for every single bug with the right test? Well, this would have been impossible before, but by giving it the goal, I can just let Codex pursue that goal and it will continue until it's done. And in this case, it's probably going to find like 24 25 bugs from the linear backlog and start crushing them as we go.

[00:19:13]And maybe like one last fun thing to show you um is the ability to create like some fun stuff with codeex using imagen. So we have an inab browser in this case and I just ask uh image and hey could you make a game that's like you know the polar bear from snowflake trying to catch some snowflakes and that's kind of like what the first concept looked like and then I can actually visualize in the inab browser directly all of these things. And what's really cool is that from here I can just talk to Codex to make changes to this game. So here I can say, "Hey, could you implement a wireless Xbox controller to control the bear?"

[00:19:51]And just like that, I can send the task.

[00:19:53]I can actually play the game while the game is being built. I think the hardest challenge will not be the AI. The hardest challenge will be the Bluetooth.

[00:20:00]It takes longer than to write the code these days. All right, let's see. Let's see it. There we go. I can just now play with my Xbox controller. And that took me five seconds or maybe 10 seconds to build. Uh, which is really really cool.

[00:20:11]Um, I'm starting to run out of time, but this was a quick tour of codeex and how all of these tools from the inab browser to like computer use kind of come together to um, let you build really anything. So with that, uh, let's quickly go back to slides to wrap up.

[00:20:28]Awesome. So, if there's one thing I want you to take away from this like presentation and this quick demo is that really like the agents to do great work need more than a prompt. What they need is a complete environment. And what that means is that they need to have the right context from all of your tools and your services through the plugins. They need to have those right tools either in the cloud or locally like computer use or in that browser to review their work and execute correctly. And then they need to have everything they need to validate that work. And so for instance in the case of a goal command it will have to run some tests or in the case of the inapp browser it can take screenshots. If you're asking codeex hey make my app responsive what it will do is like it will open up a browser try to resize the screen sizes and and continue that loop until it's done. And finally it will have to have a way for you the human behind the agents to review that work and and give it feedback in order to continue.

[00:21:24]So with that, I hope this gives you some ideas on how you can raise the ambition of what you can build and what you can accomplish with agents. And we can't wait to see what you build next. Thank you so much.