The Model Doesn't Matter. The Harness Does.

The model doesn’t matter. The harness does.

This is a follow-up to Let’s Talk About “Claude”. In that post I tried to split “Claude” into two parts, the harness (the experience on top of the model) and the models themselves. For most of us, when we say we love Claude, we’re talking about the harness. I still think that. I live in these tools, and I’ve landed on a blunter version anyway.

The model doesn’t matter. Everything since Opus 4.8-ish has been good enough for day-to-day tasks outside of the hardest engineering and cyber problems. That exception is real. The hardest engineering, and the hardest cyber problems, are a different job. For the day-to-day stuff, I don’t think swapping models is where the time goes anymore.

The agent harness is what matters: context, connectors, heartbeats, browser use, etc.

Here’s the version of that I use. Muse + Instinct + GrokBot + Dot for personal work. incident.io for production reliability. Optimizely and Profound for marketing. Cognition, Claude Code and SpaceX Super Intelligence for engineering. Etc.

None of that is a model ranking. It’s a list of harnesses, one per job, because personal work and reliability and marketing and engineering don’t want the same context or the same connectors. If the context isn’t in there, you get writing that doesn’t know your company. If it can’t connect to the tools, you’re the one copying things in. If it only runs when you remember to open it, you don’t really have a heartbeat. And if it can’t use the browser, a lot of tasks stop at a draft you still have to go finish.

I think the leverage point these days is taking full advantage of the harness.

Originally posted on LinkedIn.