A couple of months ago I would have told you nothing could move me off Claude. Three weeks ago I moved the bulk of my work to Codex.
The reason has nothing to do with Opus 4.8 versus 5.5. It is the harness.
By harness, I mean everything around the model that turns it from a clever respondent into something you can delegate work to: persistent context, access to the right files, clear role instructions, permission boundaries, handoffs between tasks, review loops, and the ability to act inside a real workspace rather than sit in a chat thread.
So I built an agent bench: a small group of agents with distinct roles across Product, Strategy, Delivery, Research, Marketing, Content, and Chief of Staff. Each has its own instructions, source rules, and handoff expectations, stored in files instead of buried in chat history.
I am not saying you cannot do the same with Claude or other tools. But the way it works in Codex fits nicely with a mental model we are all accustomed to: a team with distinct roles and responsibilities.
I talk Product with the Product agent, Strategy with the Strategy agent, Research with the Research agent, and Content with the Content agent. Each has one constant, always-on thread, so the work does not collapse into one muddled mega-thread.
It also works beautifully from my phone, which matters more than I expected. I can check progress, steer a task, or unblock an agent on the move, which makes it feel less like a chat session and more like managing work already in motion.
I have also started giving the system real operational context. The Chief of Staff now has access to my calendar and email, along with very clear rules of engagement. Email stays read-only for now, but even that changes the role. A Chief of Staff agent is much more useful when it can see the week and surface what needs my attention.
The point is not autonomy for its own sake. I still want the important decisions coming back to me. But there is now a useful amount of movement inside the system before that happens.
The clearest example is the loop between product and engineering. A product agent can take a rough idea and turn it into a proper spec. An engineering agent can pick that up and build against it. The product agent can review the result, push back where the implementation has drifted, and keep the thread moving until there is a real decision that needs me.
That is the part that feels different. The work is no longer waiting for me to manually carry every piece of context from one stage to the next.
The old constraint was how much I could hold in my own head, and how quickly I could move it through the business. That constraint has changed, not because AI replaces founder judgement, but because it gives the work somewhere to go before it comes back to me.
For me, Codex has started to feel less like a smarter chatbot, and more like an operating layer that I can delegate real work to with confidence.