The Codex Desktop app dropped quietly, as these things tend to, and then very loudly in every feed I was watching. The desktop interface (for OpenAI’s new ChatGPT 5.5 model) was also a new place to work. The reactions were different from the usual noise. The teams I'm paying attention to are moving from Claude to Codex.

(A quick clarification before the tool fight starts: I am talking about the desktop apps here, not the terminal or CLI versions.)

My working context stack lives in the Claude ecosystem. I have built custom skills and a personal constitution that I trust enough to run client work through. When people I respect move toward something new that fast, it is worth paying attention.

So I gave Codex a real try.

There is a framework I have been using: the Relationship Model. The premise is simple: every AI task falls into one of two modes. You are either writing a brief or having a conversation.

Briefing the AI means you know the outcome, you describe it clearly, approve the plan, and step away. Conversation mode means you are in the room while it happens, shaping the direction, correcting in real time, building toward something you could not fully spec out in advance.

These feel like the same activity because they both start with you typing something, but they aren’t. And the tool you reach for should match the relationship the work actually requires.

Claude Cowork (originally built for knowledge workers) is a briefing tool, and I mean that as a description of its design philosophy. The AI runs inside a sandboxed virtual machine with the specific permissions you have granted it, which is the architectural version of “you can use the studio, but check in before you rearrange the furniture.” You describe the outcome, review the plan, approve it, and go do something else. Give it a folder of source material, ask it to produce a client-ready synthesis and let it run. This is work that often gets worse when you hover over it. Time away from Cowork made that clearer: when the task calls for a brief, it is a very good collaborator.

Claude Code (though originally built for coders) becomes useful when the work is less about handing off an output and more about navigating a system. It can inspect files, run commands, test changes, debug errors, and respond to what it finds. That makes it especially strong for prototypes, codebases, small tools, technical troubleshooting, and any project where the next step depends on the result of the last one.

Claude Desktop already contains its own version of this split, and understanding the choice between Cowork and Claude Code is where the model gets practical. Cowork is for work you can fully describe before it begins. You know what done looks like, you write the brief, and the sandboxed environment runs without you hovering. Claude Code is for work you have to discover in motion: a landing page that needs ten rounds of iteration or a project where each correction changes the next step. The permission model is wider, the access is deeper, and you are present for every move.

Most strategy and creative work calls for Cowork most of the time. Claude Code earns its place when the work requires you to be in the room, directing in real time, not waiting for output.

OpenAI Codex answers that same desktop-app question differently. Open it and the spatial logic changes entirely. Your conversation runs on one side and what the agent is actively building sits on the other: a live browser, a rendered interface, a terminal, images generated and folded back into the work in the same frame.

Brief mode and conversation mode are both available, but you do not leave the room to switch between them. You can watch the agent navigate a real website, test an interface it just built, catch an error, and correct it without changing applications. Multiple agent threads can run in parallel, each with its own environment, while you move between them from a single sidebar.

For anyone already working in the Claude ecosystem, there is an additional twist: Codex can run Claude Code inside its terminal window.

I have been building inside Claude long enough that I now have an ecosystem, not just a tool. That ecosystem is the accumulated result of many conversations that have become infrastructure, which can now be portable.

So before the tool selection, the mode selection.

Look at the week you actually had. What did you produce? How much of it required you to be present while the AI worked versus just receiving the output? Do you regularly cross between knowledge work and building inside the same session, or do those activities live in different parts of your week? The answers to those questions do most of the work.

If your work is mostly documents, research synthesis, structured outputs, and operational tasks, and you are already on Claude, your stack may already fit.

If your work constantly crosses between thinking and building, and you want a single window where the agent can move from planning to execution to testing without losing the thread, the unified model in Codex starts to make real sense.

The point is not to know the shiniest new tool; it is knowing which ecosystem aligns with how you actually work. If you are evaluating Claude, Codex, or the broader agentic tool stack, start with the work itself: what can be briefed, what needs to be steered, and where your current workflow keeps breaking down.

I help teams make those decisions, then turn them into systems people can actually use. Reach out if you are ready to choose the right AI stack and build a workflow that fits the way your team actually works.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

I’ve also started publishing more frequently on LinkedIn, and you can follow me here

if you’d like to chat further about opportunities or interest in AI, or this newsletter, please feel free to reply.

banner images created with Midjourney.