Both big labs now sell delivery as well as models, and that’s a reasonable response to a real problem: most enterprises are still struggling to get measurable value from AI, and the labs want to show it can be done. Just be clear about what you’re buying. A lab’s consultancy will teach you to use its own models and tools very well, and it has no incentive to build something you could easily move to a competitor. The models themselves are fairly interchangeable now. The harness around them, meaning your agents, workflows, memory and permissions, usually isn’t, and that’s where lock-in sits, whether anyone intends it or not. If a lab-backed team is building for you, I’d ask them early on to show how the work would move to another provider’s model and what that would cost.

I’ve kept an eye on duplex voice models for a while, and Dessa’s demos are impressive. Most voice agents today chain three models together (speech to text, a language model, then text back to speech), with a lot of engineering to make interruptions feel natural. Dessa does all of that in one model, which is why it laughs, interrupts and says “yeah” in the right places. The catch is control. Duplex models are bad at waiting, so they call tools at the wrong moment, and they’re hard to steer with instructions. Even OpenAI’s GPT-Live doesn’t let its voice model call tools. It hands that work to a separate text model in the background, which sends back results and guidance. When Block, or anyone else, releases an API with dependable tool calling and published benchmarks, we’ll test it against today’s pipelines. Until then, clients should keep building on what they can govern.

Duet is the latest in a run of proactive personal agents from the major labs, after Meta’s Muse and OpenAI’s Dots, and they’re all getting closer to being genuinely useful. What interests me is where these agents sit. At the moment you go off and ask one something, or it pings your phone with a briefing. The next step is the agent in the meeting with you, following the conversation and contributing as it happens. I think that’s when enterprises will really see what agents can do, because the value shows up in front of the whole team rather than on one person’s screen. It also raises questions most organisations haven’t answered yet. We're used to our calls being transcribed as default, but agents can act as well as listen. We need to be thinking about our agent governance today because it won't be long till they start appearing in our meetings and channels.

We’re big advocates of ontologies, so it’s good to see an engineering team as respected as Netflix’s show one working at this scale. The post shows, clearly and in detail, the pattern applied to a problem almost every large organisation has: signals from dozens of systems that nobody can join up when something goes wrong. Netflix mapped that noise into one shared model of how its users, devices, services and infrastructure relate, then let agents reason over it to find the root cause. That’s why ontologies matter more now. Language models are only as good as the context you give them, and an ontology is the most dependable way to give an agent context it can trust across an organisation. If you have a problem like this, where the data is noisy and spread across teams, I’d start there. Build the semantic layer for that one job and let it prove its value before you extend it.