When AI Agents Start Buying and Selling Alone

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

Right. Anthropic just ran an experiment that's either fascinating or deeply unsettling, depending on how much you like the idea of software negotiating on your behalf. They built a marketplace where AI agents bought and sold things from each other. No humans involved. Real goods, real money, real deals. Let's talk about what happened.

So here's the setup. Anthropic created what they're calling Project Deal — essentially a test version of Craigslist, but every buyer and every seller was a Claude agent. The agents were given budgets, told what to buy or sell, and left to haggle. And they did. Successfully. Cameras changed hands, furniture got traded, prices got negotiated down. All agent-to-agent, no human in the loop.

Now, we covered this briefly last week, but the details that have emerged since are the interesting bit. Because it wasn't just that the deals happened. It's that they happened smoothly. Agents adapted their strategies mid-negotiation. They made counteroffers. They walked away when the price wasn't right. In other words, they behaved like people. Which raises a question nobody's quite answered yet: if two agents agree to a deal and one of them gets a raw deal, who's liable? The agent? The person who deployed it? The company that made the model?

And here's why that matters. If this works at scale — and Anthropic clearly thinks it will — you're looking at a future where chunks of commerce happen autonomously. Procurement agents negotiating with supplier agents. Marketplace algorithms that don't just recommend prices, they set them and agree to them. That's efficient, sure. It's also a legal minefield. Competition law wasn't written with algorithmic collusion in mind. Contract law assumes humans read the terms. We're about to find out how well those frameworks bend.

Right, moving on. A German robotics company called Sereact just raised $110 million, and if you haven't heard of them, you're not alone. They're based in Stuttgart, they've been quiet, and they've just become one of the better-funded robotics software companies in Europe. What they're building is software that lets industrial robots handle tasks they haven't been explicitly trained on. So instead of programming a robot to pick up part A and place it in bin B, you give it a general understanding of objects and let it figure out the rest.

This is the embodied AI trend we've been tracking — robots that don't just follow scripts, they adapt. And the funding round, led by Headline, suggests investors think the market's ready. Because the alternative — training a robot for every single variation of every single task — doesn't scale. If you're running a warehouse or a factory and your product mix changes weekly, you need robots that can learn on the job, not ones that need a software update every time you stock a new SKU.

What to watch here: Sereact's competing in a space that's heating up fast. You've got Google's embodied reasoning models, Tesla's Optimus, a dozen well-funded startups in the same lane. The question isn't whether this tech works — it clearly does. The question is who builds the stack that enterprises actually adopt at scale. And right now, that's still wide open.

Now, one last one, and it's a supply chain story that matters more than it sounds. TSMC and Nvidia just celebrated the first Blackwell chip wafer manufactured in the United States. Not Taiwan. Arizona. And if you're wondering why that's a headline, it's because this is the first time in recent memory that the most advanced AI chip in the world has been made on American soil.

Nvidia's committed to producing up to half a trillion dollars — yes, with a T — of AI infrastructure in the US over the next four years. They're partnering with TSMC's Arizona fabs, and they've locked in the majority of TSMC's most advanced packaging capacity. Packaging, by the way, is the step that connects the chip to the board. Unglamorous, essential, and increasingly the bottleneck in AI hardware production.

Why does this matter? Two reasons. One, it shortens the supply chain for the infrastructure underpinning most of the AI you use. If a geopolitical event disrupts TSMC's Taiwan operations, US-based production is a hedge. Two, it's a signal that the US government's industrial policy is working. The CHIPS Act put serious money into reshoring semiconductor manufacturing, and this is the first genuinely flagship product to come out the other side.

Keep an eye on yield rates and timelines. TSMC's Arizona fabs are new. Ramping up production of something this complex, at this scale, is hard. If it works, it resets the map for AI infrastructure. If it doesn't, Nvidia's got a very expensive problem.

That's your lot. Three stories: AI agents trading without supervision, robots learning to improvise, and the world's most important chip now being made in America. If any of that was useful, tell someone. If not, well, you know who to blame. See you next time.