Claude Duet Works While You’re Away—Who’s Watching?

James Turner

Here's a small detail from Anthropic's launch announcement that tells you what Claude Duet is. A broker named Jess Provencher gets a rundown from it every morning. One day, without being asked, it flagged a property showing on her calendar that was on the wrong day. Nobody typed a prompt. The software noticed and came to her. That's the whole pitch, and it's why this is the lead story in the 48 Hour AI roundup for October 21 to 23.

James Turner

Anthropic opened a research preview of Claude Duet on October 22. It's a standalone app on iOS, Android, and claude.ai, built around Claude Opus 5.5. Here's the mechanism. Duet runs in its own cloud virtual machine with its own browser. So your laptop or phone can be off and the work keeps going. It sends a morning brief across your connected apps, and it checks those accounts every few hours, not constantly, because Anthropic says constant checking would burn through your plan limits much faster. It plugs into more than 6,000 connectors in the Claude Marketplace, things like Gmail, Slack, QuickBooks, Stripe, Gusto payroll, HubSpot, Docusign, and Shopify.

James Turner

Now think back to our October 21 episode, where agent sessions were isolated silos that evaporated when you closed the terminal. Duet is the opposite design. It's persistent, it carries over your Claude memory, and it keeps one ongoing thread. Which raises the question from that same episode: who's watching it?

James Turner

Anthropic's answer is layers. Duet never directly sees your passwords or sign in tokens. Actions go through the same auto mode classifier that guards Claude Code. There's a new classifier built to reject requests to share the strictest tier of information, like credit card numbers and government ID. Health or personal finance details only go out if you explicitly told it to send them. And before it sends a message, makes a payment, signs a document, or deletes anything, it shows you the action and waits. Some things are off limits even with your approval, like trading securities or moving money between accounts. For purchases it uses Stripe's Link with a virtual card, so your real card number isn't shared.

James Turner

Anthropic is also candid about the limits. It might misread a thread or miscalibrate what's urgent. Personal errands like booking tickets are weaker, and it can't yet show you its browser. That last one matters. The agent works in a window you can't see into.

James Turner

Which brings us to METR. On October 21 it published a review of the misalignment section of Anthropic's own risk report. METR largely agrees that risk from the covered models is low, but says that's because of limited capabilities, not robust safeguards. It also warns that current evaluations may not reliably catch risk as capability grows, and that the benchmarks don't test whether a model can deceive Anthropic's monitoring. And it points to recent incidents that revealed real world misalignment nobody had detected.

James Turner

One number from a newsletter summary of the review, The Signal, stuck with me. Anthropic's offline monitor flagged about 1% of one model's actions. With the chain of thought stripped out, so the monitor only saw user turns, tool calls, and tool results, that jumped to about 50%. That's the newsletter's summary, so check the review itself. But if it's right, a lot of the monitor's confidence came from the model narrating its own work. Now put that next to a background agent whose browser you can't watch. Approval gates help, but they only cover the actions someone thought to gate.That's the tension to carry into everything else today.

James Turner

Okay, the rest of the week, starting with something that runs entirely on your own machine. Recursive released Flint, a mixture of experts coding model with 40 billion total parameters that activates just 2.5 billion per token. It was built from scratch by a small team of people and agents in under four months. Reported speeds: more than 125 tokens per second on an M4 Pro MacBook Pro, and more than 150 on a 64 gigabyte NVIDIA RTX Spark at short contexts. The pitch is a daily coding model that finishes tasks locally without sending your code to a remote server. Those are the developer's numbers, so test it on your own repo.

James Turner

On the hardware side, Unitree unveiled the H2 X Kung Fu Legend Guardian, a humanoid for security and rescue. It's 182 centimeters tall, weighs 70 kilograms, and has 31 movable joints. Unitree says it runs eight hours on a charge, nearly triple the base H2's three hours, with 2,070 TOPS of onboard compute. It reportedly executes martial arts moves using a model that predicts opponent actions in real time with no human controlling it. It can jump off four meter platforms and lift cars off trapped people. And Unitree itself warns people to stay two to three meters away while it operates. No price or release date yet. For reference, the base H2 starts at $29,900.

James Turner

In infrastructure, Google and Meta brought native PyTorch support to TPUs with TorchTPU. The idea is that a developer changes their device initialization to tpu and runs the same training loop. Google's engineering write up describes three eager modes, and the fastest, Fused Eager, fuses operations on the fly and delivers 50% to over 100% better performance than Strict Eager. For full speed it compiles through torch.compile and XLA. My read, and this is interpretation, is that easier migration between GPUs and TPUs puts pressure on NVIDIA's software advantage. But Google itself says portability doesn't erase hardware realities, and models may need refactoring to use TPUs well.

James Turner

OpenAI announced it's acquiring Orchestra Bio, a platform that coordinates drug development for humans and AI agents working the same program. It'll be combined with Rosalind, OpenAI's life sciences stack. Terms weren't disclosed. OpenAI's example is a failed manufacturing batch that delays a trial, where the agents work out which downstream steps are affected.

James Turner

And now the money. The Wall Street Journal viewed a preliminary Anthropic S1, and the reported numbers are striking. First half 2026 revenue of $16.3 billion, up from $1.1 billion a year earlier, roughly fifteenfold. Third quarter revenue reportedly hit $16.8 billion, with a November listing targeted at a valuation reported as high as $2 trillion. Cloud commitments total roughly $450 billion over a decade. The detail I'd underline is in the risk language. Anthropic says it has trouble securing capacity from AWS, Google, and Microsoft, which increasingly treat it as a competitor, and that the other data center operators it's turning to have a limited operating history.

James Turner

Meanwhile, Nvidia's quarterly filing discloses up to $164.5 billion in guarantees and other commitments behind data centers that run its chips. About $105 billion of that, according to The Information's breakdown, is a guarantee on lease payments for the Ohio campus SB Energy is building for OpenAI. If OpenAI doesn't pay, Nvidia does.

James Turner

So here's the thread running through all of it. Demand for agents is compounding, the chips and leases behind them are being financed in unusual ways, and the tools to watch what those agents do are still catching up. Duet is working while your phone is off. Whether anyone can see well enough into its work is the open question. That's the 48 Hour AI roundup. I'm James Turner, and I'll see you next time.