The Harness Wars Have Begun

First, let's define the buzzword. "Harness" is tech's latest term for the layer that sits between you and an AI model. The thing that lets an AI actually do work on your computer instead of just talk about it. It connects the model to your files, your apps, your browser, your email. Without a harness, AI is a chatbot. With one, it's a coworker that can open your spreadsheet, draft the email, file the document, and set up your software. Claude Code, Cowork, Codex. These are harnesses. The model is the brain. The harness is the hands.

Yes, it's a buzzword. I'm using it anyway because there isn't a better word yet, and you're going to start hearing it everywhere.

Three months ago, this looked like a one-horse race. Anthropic had Claude Code, the first harness that actually worked for real development. Then they launched Cowork, which took the same power and handed it to people who don't write code for a living. Business owners. Operators. People like my clients.

For months, Anthropic owned this space. Nobody was close.

That's over now.

OpenAI completely rebuilt Codex. Not a refresh. A rebuild. And last week they dropped GPT 5.5, a model that has moved ahead of Claude Code and Cowork in ways I didn't expect this fast.

It's faster. Noticeably faster. The coding output is better. And the workflow design shows they've been paying attention to every friction point Claude Code created and solving them.

I switched 85% of my workload to Codex last week.

I'm not saying that to be dramatic. I'm saying it because it's true and you need to hear it if you're building your operations around these tools.

Beyond speed, the permission model is the next big one. If you've used Claude Code or Cowork, you know the constant interruptions. "Can I access this file?" "Can I run this command?" "Can I open this application?" Every permission request breaks your flow.

Codex auto-accepts most permissions and only stops you on the ones that actually matter. The high-stakes stuff, like deleting data or accessing sensitive systems, it pauses and asks. Everything else just flows.

That sounds like a small thing. It's not. It's the difference between a tool that works with you and one that works near you.

I just had Codex use computer control and the browser to set up my new company's Microsoft 365 connector app. Think about that for a second. Instead of navigating Microsoft's labyrinth of admin portals, clicking through fifteen screens of Azure AD configuration, I described what I needed and it did it while I read some emails. You don't need to go into old software interfaces anymore. The harness is the interface.

Here's the honest version: I jump back and forth. I stay in Codex for the more important work right now, but I keep finding myself back in Cowork for specific things. I wrote this post there, for example. The connector ecosystem Anthropic has built is deeper. Some of my workflows are tuned to how Claude thinks. And there are things Cowork does with document creation and multi-step business tasks that Codex hasn't matched yet.

Different tools are better at different things, and right now neither one has pulled away enough to be the only answer.

Here's what makes this a real war and not just an OpenAI comeback story: Anthropic is shipping at a furious pace. New features are landing weekly. The platform is evolving fast enough that the gap Codex opened could close in a month.

And the distribution play is massive. Microsoft is licensing Cowork. Let that sink in. The company that owns the enterprise desktop is putting Anthropic's harness in front of every Office 365 user on the planet. That's not a product advantage. That's a distribution nuclear weapon.

Then there's Mythos. Anthropic's next model is in pre-release right now with a handful of key customers. They're patching every bug it finds, and early reports suggest it could be a meaningful leap. Is it ahead of GPT 5.5? I don't know yet. Nobody outside those early testers does. But the fact that it exists and it's close means this race is far from decided.

If you're a business owner building on these tools (and you should be), here's the reality:

Earlier this year, it looked like Anthropic was pulling away. They had the best model, the best harness, and no real competition. You could bet on Claude and feel confident.

That's not the world we're in anymore. OpenAI caught back up and moved past them a bit. Anthropic has the distribution deal and a new model in the chamber. Both platforms are shipping features at a pace that makes last year look like a nap.

The harness wars have begun. And unlike the model wars, where one company pulls ahead for a few months until the other catches up, the harness wars are about ecosystems, workflows, and user trust. Those are harder to build and harder to lose.

My advice: learn both. Build on both. Don't marry a platform right now. The winner of the harness wars won't be decided by a benchmark score. It'll be decided by which tool you reach for first on a Monday morning when real work needs to get done.

Right now, I'm reaching for Codex more often than I expected. Ask me again in a month.

Alan

PS I also think Google has fallen behind for now. It was the leader in images but look at the thumbnail for this post. OpenAI shipped a new image model with 5.5 and it rocks.