Always-on agents. The model is swappable, the harness is not.

A product person at OpenAI’s DevDay calls her new agent on the phone, the way you would call an assistant, and asks it to do something. The agent says it’s on it. Then about ten seconds of silence. She asks for an update, and it says it’s still checking. She jokes that it’s having a “slow morning”. That was the launch of Dots, OpenAI’s always-on agents, and the clip went around for days.

I got asked about it a lot in the last week, so I want to write down my take in one place. The demo stumbled, and they blamed the Wi-Fi. Fine. But I don’t think the stumble is the story. I think Dots is the most honest product OpenAI has shipped in a while, because it admits what an agent actually is.

An agent is a process

An agent isn’t an application sitting in a stable state, waiting for you to type something. It has state. It has a loop. It keeps going after you close the tab. By definition it’s a process that runs in the background, and for two years we’ve been squeezing that process into a chat window and pretending the conversation was the product. If you’ve read the agent loop post, this is the same loop, just with nobody watching it turn.

Dots drops the pretence. Per OpenAI’s own description, each Dot runs on GPT-6 Astra with its own cloud computer and browser, connects to apps through plugins, and keeps working between conversations, with rules for when it can act alone and when it needs your approval (technology.org). That’s a process with permissions, not a chatbot with memory. So it’s the natural evolution of an AI application, and I’d bet a lot on it. We’ve already seen the same idea coming from Grok and from projects like OpenClaw. I expect this to become a new form of app, something that lives on your iPhone next to everything else, within a few years.

As for the demo: I think they’re moving very fast. A lot is going through the pipeline at once, and not everything comes out the other end at the quality you’d want on a stage. That’s a cost of the pace, not a flaw in the idea.

The moat moved to the architecture

So is the model still the moat? No. It’s the agent architecture, and I don’t think that’s close. The thing worth building toward is an agent harnessed so it can run in the background with its own knowledge base and its own tasks. That’s what Dots is reaching for, and it’s the part that’s hard to copy, because none of it comes in the model weights.

The model race is getting harder, not faster

The same week Dots launched, OpenAI cancelled GPT-6.1 Astra, an agentic model planned for ChatGPT and Codex, because internal safety testing found it overstepping its authorisation and misreporting the work it had done. They kept pushing agents anyway, on the existing GPT-6 Astra. I read a lot into that.

We’ve had a run of models pushed out on very tight schedules. Competitors release on the same day. Grok has done it, Anthropic has done it, Chinese labs are trying to keep the same cadence. It’s a brutal space to compete in, and every new model is harder to ship than the last one, because the bar keeps rising and so does the cost of getting it wrong. Will we keep seeing significantly more frontier releases at this pace? I think that becomes less likely, not more.

A cartridge for AI

The news I kept coming back to wasn’t from OpenAI at all. In August AMD agreed to acquire Taalas, a Toronto startup that builds inference chips with a specific model’s weights hard-wired into the silicon instead of loaded from memory. Their first demonstrator runs Llama 3.1 8B. The model and the chip become one object.

That’s basically a Nintendo cartridge for AI. You don’t download the model, you plug it in. I’d bet we see something shaped like that within the next couple of years, and it says a lot about where the money is going. Improving the model itself is becoming less important than improving the infrastructure that runs it. Capability costs the consumer money. What the consumer actually cares about is the application: what they can do with it, whether it’s cheaper, whether it’s good enough to reach everyone.

So expect more infrastructure releases, and expect them to be copied fast. Something ships, and within weeks Amazon or someone else has their own version. That pattern is going to get more common, not less.

Distribution is the real moat

People ask me whether workflow integrations are a moat or just a distribution channel. I think they’re the real moat. Every integration that actually gets used is proof the agent is useful, and that proof is what lets you raise usage limits and raise what an enterprise is willing to spend. That’s my hypothesis, anyway, and so far the evidence keeps agreeing with it.

We’ve been through a lot of hype cycles about generating applications. Sometimes it works, sometimes it doesn’t. What we need now is to push into real work where we don’t necessarily need to generate code at all, and use AI as infrastructure for running the company in the background. Reconciling the invoices, chasing the follow-up, keeping the CRM honest. Nobody needs a new app for that. They need a process that does it.

What I tell a CEO

If a CEO asks me whether to care more about model choice or agent-system design, here is the one-paragraph answer. We’ve seen tiny models with small context windows and giant ones with a million tokens. We’ve seen models that reason well and models that win benchmarks in one domain and lose in another. And yet almost all of them sit behind pretty much the same interface. You can make nearly any of them OpenAI-API compatible and swap one for another in an afternoon.

The hard part is everything around the model. It depends on your proprietary data, your process knowledge, your integrations with internal systems, all the things that are specific to your company. Building that agentic system is far more difficult and nuanced than swapping in a newer, more capable model. And LLM generation on its own isn’t as reliable as an agentic system needs to be. A lot of the system has to be built with the same boring software engineering we were doing five years ago: queues, retries, permissions, tests, verification.