Codex vs Claude Code: Stop Picking One. Here's What I Do Instead.

0:00 Everyone is asking whether Claude Coder or Codex is better. I I literally get this question. Nate, you talk about Codex, does it mean you don't like Claude Code? Nate, what do you think about Claude Code? Can I do this in Claude or Codex? Those are the wrong questions. The better question is what does each tool make you better at doing with agents? Because the skill of 2026 is agent literacy. And I'm going to give you a shorthand at the top here. I think Claude makes steering agents feel very, very natural, and Codex makes dispatching agents feel very, very natural. We're going to get into those differences in this video. That difference may matter more than which model wins a benchmark this month, because it teaches you a habit. Look, this is like the Mac versus Windows fight of the agent age. Not because Claude is Mac and Codex is Windows or the other way around. That's too cute.

0:43 The point is that interfaces train behavior. Mac and Windows did not just compete on features. They actually taught people what a computer was for, where work lived, how files moved, how much the machine should hide or show, how much control the user ought to have. So, Claude and Codex are doing that now for agents. They are teaching us what an agent is for, and that is why this matters even if you don't write code. The names make this sound like a developer fight, right? And a lot of developers use these tools. So, Claude Code, Codex, work trees, hooks, sandboxes, diffs, you hear these words and you're like, what are they? And I get why people are like, these tools are not for me. But I think that this is one of the first AI debates that non-technical people should force themselves into the room and say, "No, we deserve to understand this." Because coding agents are where agent habits we all will use are showing up first. A chatbot answers, and an agent takes a job, right? That's the simplest distinction. That latter piece, the agent taking the job, that's the piece we have to get fluent at directing. And so, you have to be able to say, "Here's a folder, here's a goal, here's what done means to your agent, and here's what you're allowed to touch." And the agent will then read files and call tools and open pages and run commands and edit drafts and check what happened and come back with something you can check.

2:00 That showed up first in coding because code has proof that of what good looks like built into it. Does the code run or does it not? Most knowledge work was not that easy and so that's why all of these tools showed up as coding tools first. And now knowledge work is coming up because these agents are getting better and that's why Claude versus Codex and understanding their different approaches to agents matter now. So, the coding world is giving us the vocabulary that both of these tools run on and I want to translate that very, very quickly. Once you translate terms like this, this entire tool set becomes way less intimidating. These are just the parts of a serious assignment, right? You need to have context and permissions and tools and checkpoints and helpers and proof if you're doing real work. Now, the Claude versus Codex question gets really interesting. Claude code feels like a cockpit that you're flying, right? When in an airplane. You're close to the model, you're steering the model, you're talking through the work while it happens. You can ask it to read the code base or the source folder and tell you what is going on. You can ask it to interview you before writing the spec.

3:04 You can stop it, you can correct it, you can make it rethink the plan, you can keep the work really close. And that feels like a real advantage when the work is fuzzy. You want Claude to be close to you, right? This is the experience I've had with Claude co-work, it's the experience I've had with Claude code. Is it subjective? Yes, but a lot of folks agree with me. If the hard part is taste, if it's ambiguity, if it's design judgment where you want to get close to the problem and really wrestle with it, if it's writing, if it's if it's architecture or figuring out the actual question, Claude is really, really good at that. The personality matters there and that that that can sound really soft, but it's not. If a tool feels patient and it feels thoughtful and it feels focused on the right solution, you can bring it a half-formed version of the problem and you can bring it something you you quite name yet and you can figure it out together. Serious Claude code users, serious Claude co-work users are not just chatting. They use plan mode before edits. They keep a Claude.markdown file, which is basically a standing note that says, "Here's how the project works.

4:06 Here are the commands. Here are the rules." They use hooks so that important checks run automatically. They use MCP servers to connect tools. They split work across sessions. A session can write and review and investigate and test. That's real agent work. The risk is that you're assembling a lot of the system yourself. You are managing the context window more. You are deciding when it makes sense to do a planning session. You are deciding how to handle hooks when you want to put hooks into your system to do automated reviews.

4:34 You're thinking about when to invoke workflow mode, which is a brand new mode in Claude that lets it spin out sub-agents. and so, if you're very disciplined, that's an incredibly powerful tool because you have all of these tools lined up in front of you and you can use them to get close to the work and really drive a lengthy work session productively. But, if you're not careful, the conversation can become a bit of a junk drawer. The context can fill up, which is a bigger risk with Claude right now. Codex feels different.

5:04 Codex feels more like an operations desk. I can have one thread reading a folder, another drafting a document, another checking a package, another using the browser, another turning a repeated process into a skill all at the same time. There's a lot more parallel compute with Codex right now because the Anthropic team is still looking for compute, right? The work queue is visible with Codex. The jobs stay separated. The outputs are inspectable very easily. And that changes what I'm willing to hand over.

5:30 With Codex, I still ask for help thinking, but much more often I say, "Go do this piece of work, bring back the results, and show me the proof you did it." For software, that could be a diff or a test output or a PR. For knowledge work, the proof might be a source list or a rendered document or a comparison table, or even just a doc that summarizes what happened, and then the source docs that show that actually that got done.

5:53 And that's why Codex feels a lot bigger than coding to me. OpenAI started Codex in software because software has very clean feedback loops. And that's actually the same reason Anthropic started Claude code on software. But the shape of work, of course, has gotten broader, and that's why these larger conversations are needed because you can use the same workflow of assigning a task and setting a goal and using tools to do a lot of other knowledge work now.

6:15 A sandbox just means the agent has a contained place to work. It can try things without touching everything else, and it can use tools. It can use skills. It can work on work that's separated out in a work tree, all without touching the rest of your machine. And this has made Codex feel really safe to use, especially now that the auto review means that there's a separate 5.5 Codex model that checks what my execution 5.5 model wants to do in Codex and make sure it's aligned with my intent before it lets it do stuff.

6:50 And that gives me context to sometimes go outside the sandbox with Codex. And going outside the sandbox means computer use. As in letting Codex take control of my computer, which is something that is much more flexible with Codex than with Claude right now. Computer use means that Codex can see and click and type on the screen, and I don't even have to be there. Right? There are background automations that mean Codex can wake up and run later and do work without me having to pay attention to it.

7:16 So, this is not just a feature list when you stack all of these together. This is a way of making agent labor easy to manage. And that's why I'm loving Codex so much right now. Not because Claude is weak, it's actually an incredibly strong model. 4.8 is really good. Claude is one of the most important AI products in the world, and Claude code has pushed the whole category forward, and there's a reason so many developers use it. But my bottleneck is often not can I think about the work? My bottleneck is often moving the work across the computer rapidly, finding the file and reading the transcript and using the source and rendering the docx and using a site and copying the file to the handoff location and verifying it exists. That's not a code task. It's work on the computer and Codex has made me more willing to hand that work to the machine.

8:02 Not blindly, right? I don't trust the agent just because it sounds confident. I trust the receipts. I I make it show me the files. I make it show me the logs. And And that's the Codex advantage. It makes the assignment design feel so natural. It makes delegation of work to agents feel so natural. But Codex has a failure mode, too. A completed run can make the work feel more done than it really is. The agent will come back and it will say the task is complete and on the surface it has all the right signals of progress. But maybe it followed the instruction too pedantically. Maybe it optimized for completeness instead of quality. Maybe it used the wrong source.

8:40 Maybe it created a pile of work that now takes longer to review than it would have taken to do the little task myself. So, Codex is not perfect. And I want to be really honest about the differences and what makes them feel risky to use because they're changing the way you think about completeness and quality. So, if you're trying to learn failure modes, be careful which failure mode you're learning depending on which tool you use. Claude can seduce you with a great conversation, make you feel closer to the work than you are. Codex can persuade you that a workflow is completed when it's really not. Both still require judgment. Both still require proof.

9:17 And so, if you're trying to figure out like what to use or when to use it, let me give you a practical decision rule. Use Claude when the problem needs conversation before it can become an assignment. Use Claude when taste and ambiguity and design judgment and writing and architecture when the shape of the question is the hard part. Use Codex when the work can be written down and it's a job you can delegate. Use Codex when there are sources and files and tools and checks and artifacts that you can all call in.

9:44 Use Codex when parallelism matters, doing two or three things at once. Use Codex when you want a repeated task to become a durable workflow instead of just one helpful exchange. And use both when the stakes are high enough, right? Let one model plan and the other critique. Let one implement and the other review. Let one agent produce the artifact and and another inspect it against the standard. And then you decide. And that last part, that's not just like ceremonial, that's the job. You are not disappearing in this world. You are the human that moves to the part of the work that can't be skipped. You're deciding and compiling meaning and figuring out what work should exist, what good means, what risks matter, what proof counts, and when the output is ready to leave the machine.

10:26 That is why this is not just about software, right? People feel the power of these tools, but they also feel the stress of working with them. Managing agents is legitimately tiring in another way. You have to trust work you did not personally do without becoming careless. You have to stop micromanaging every step without becoming gullible. You have to let the machine run and then be ruthless about what came back. That is a new skill we're all learning. Learning when to steer, we're learning when to dispatch, we're learning when to verify.

10:55 And this is the agent literacy I care about. Yes, prompting is a part of it, but prompting is far too small a word for what we're doing here. We're doing agent loop management now. The skill is writing assignments that come back as inspected work. And the interface war that we're talking about with Claude versus Codex is over which product makes habits that feel natural with our workflows. Which one makes you ask better questions? Which one makes you write cleaner assignments? Which one makes permissions obvious? Which one makes it natural to run more than one agent? Which one makes proof hard to forget? These are human questions that come up because of the agent interface, right? Which one turns repeated work into a skill, a hook, an automation, a workflow? That's what I am watching.

11:37 Right now, I don't think the honest answer is Claude wins or Codex wins. They're pulling the future in different directions, and I'm keeping an eye on both. Claude is very good at keeping the agent close while the work is still becoming very, very clear. Codex is really good at making agent work feel very assignable and parallel and and inspectable. The best users I know are using both. And if you're asking which one has changed my own work more, I have to be honest, it was Claude first and now it's Codex. They both changed the way I work a lot, and working with both has made me better. And right now, what's special about Codex is it made me stop thinking of AI as a place where I get help and start thinking of my computer as a place where work can be delegated and checked and packaged and continued autonomously. And that's the beginning of a new kind of computer literacy. That This is how interface shifts usually feel. First, they look like a niche workflow for power users.

12:34 It's the people using BlackBerry back in '07 and '08. And then it becomes the default way serious work gets done. Now everybody's on their phone, right? So, don't reduce Claude versus Codex to a coding tool debate or even to a Mac versus Windows debate. Watch what each tool makes it easier for you to imagine. Watch very carefully what each tool makes it easier for you to forget. Watch the habits it creates in you. The most important question is not which agent is smarter, guys. The most important question is what work am I now capable of running, and what proof would make me trust it, and which of these tools helps me to do that? The thing to remember, the thing to keep in mind, is that you are on the edge of the agent revolution now. We all are. Anyone who says they've got it figured out is lying. They're lying to you. We are all figuring out together how to manage rapidly evolving agents. It's a new paradigm for computing and I'm passionate about talking about the differences between these tools because the differences are going to shape the way we imagine with agents. It's going to be different if you're a Claude user in 6 months and you have mental patterns that are Claude patterns versus Codex.

13:46 Do you know how I know that? I know developers who feel like they are switching interfaces and their brain hurts when they switch from one other because they have to think differently about how agents work. The the sandbox example is a good one. Codex runs in sandbox, Claude doesn't. That's just one example among many. So, start to think about really intentionally what kind of of work feels natural to you. In in developer terms, we call this developer ergonomics, right? Like imagine being comfortable in a working space. What helps you to run agents that get work done? Is it Codex? Is it delegating the work? Is it Claude? Is it feeling close to the work and digging in and having a conversation? And these, by the way, are are summaries. If you're relying on this and you're like, "Wow, Nate says that Claude is this and and I found Claude to be that." Well, tell me in the comments, right? Like the the point is that we want to build up the knowledge together.

14:43 I believe strongly in a cool kid philosophy for AI. In other words, you should be the one who gets to be the cool kid for the day by showing all of us the amazing work that you do with AI. So, if you have a trick with Claude or a trick with Codex or a different approach to ergonomics with these, a different approach to feeling comfortable using these agents, put it in the comments. Let's talk about it. Let's learn about it together. I think it's really, really important that we take the time to dig into what makes these interfaces distinct because I believe strongly that as agents get more powerful, they're shaping the way our minds work with AI and we got to be intentional about that.

15:21 We got to pick something that feels like it works for us and lets us do work that's meaningful for us. And that's that's very personal, right? It's around aligning the subject matter we work with, the outputs we're looking for, the quality of the model, the quality of the hardness or tool like Claude or Codex, and then finally our willingness and ability to work with that tool to get the work done, and our ability to feel comfortable with that. My whole goal with Claude versus Codex is to give you enough of a taste here that you can start to either be curious cuz you've never tried both or that you can start to jump out of your seat and say, "Nate, I've got it. Nate, you're right. Nate, you're wrong." Well, tell me that, right? Tell me the comments. Let's make this a learning activity. And then if you want to get started, I absolutely have very, very detailed get started guides for both of these today. And of course that deep dive on Codex coming Friday. So, get excited for that and I'll see you in the comments. Cheers.