When OpenAI launched dots, my first reaction was that it looked like a consumer product, with cute avatars in berets and a very polished microsite. I've since come round to thinking the marketing is the smartest part of it. Managed agents were always going to reach the workplace, and most of the ways they could have arrived were dull. Most staff at the businesses I've worked with this year have barely used an agent. A dot that messages you or picks up a call works the way they already deal with colleagues. I expect people to compare notes on what their dots have said, and that word of mouth is how many of them will end up trying an agent. There are governance questions to settle first. Each dot builds up memory about your work, and it isn't yet clear where that memory is held or what happens to it when a subscription ends. Losing it would be like a good employee leaving without a handover. Agree where that knowledge will be kept, inside your own business, before anyone switches one on.

The obvious read is that you now choose the ecosystem. The actual move is the oppostite. OpenAI is working with Microsoft to bring specialist dots under Agent 365’s governance controls. Microsoft has conceded that the agents themselves will come from several vendors, and is claiming the control plane instead. If most of your estate is Microsoft 365, that is where agent identity, permissions and audit will sit regardless of whose agent you run. Which leaves on question for your team. Where will the business knowledge that these agents depend on be stored, and who controls it? An agent whose identity, memory and organisational context all live in Microsoft 365 is not easily moved anywhere else.

Most of what went wrong sits in the supervision around the model, rather than inside it. An agent was left to run for long stretches while observing activity was not as thorough as the frontier providers would strive for. That's a solvable problem. If you run agents today with internet access, ask your team whether anyone is appropriately restricting and monitoring their access. I'd give OpenAI some credit here. Its researchers have said for a while that models don't always do what they're told, and an active area of research is improving their goal and value alignment. Whatever the vendor says, if your agent has internet access, test how an agent behaves in your own environment before it goes live.

We know that building the “digital twin” of your business is crucial to enabling capable and productive AI agents within complex enterprises, so it’s great to see more tooling (and in particular, open-source tooling) becoming available in this space. On paper, it’s pitched as portable and open. You don’t have to use AWS infrastructure to build it (though of course, it’s made easier if you do!), and the ontology it produces is described in formats which you can take elsewhere. All very promising - and worth testing that portability sooner rather than later to see just how easy it is to avoid vendor lock-in.