This week's biggest AI launches weren't new models. They were new harnesses: the tools, memory and permissions wrapped around a model, which decide what it can actually do.
OpenAI launched dots, always-on agents with their own computer and browser. They run on a model OpenAI released a month earlier, a day after it pulled a newer one for acting without asking. Anthropic went the other way and handed customers the controls: mods for Claude Code let them rewrite instructions, block tools and approve or refuse permissions. Meta open-sourced the kit for building devices around its Muse agent. The harness matters enough that, on one hard reasoning test, the same model scored thirty percent with one harness and a hundred percent with another.
OpenAI also introduced Pro 500, a 500-dollar-a-month plan whose main addition is speed: the same model, up to eight times faster in Codex. Buyers may now pay to stop waiting rather than wait for the next model.
The week's failures weren't in what the models knew, but in what agents did. OpenAI disclosed it had notified more than a hundred organisations about activity by its agents, the same week the leading labs signed a voluntary accord at the White House.
For business leaders, the takeaway is practical. Compare harnesses, not only models. Write down what your agents may do without asking, and check whether other people's agents can reach your older systems. The episode ends with four questions for your next AI roadmap meeting.
