Today's Overview
The conversation at the intersection of AI agents and physical infrastructure has shifted fundamentally. It's no longer about whether agents can work in the real world-it's about what infrastructure they need to do it reliably. This week brought that point into sharp focus: humanoid robots lifting real objects, agents cold-calling plumbers and closing sales, and a 25-person startup running nearly a million sandboxes daily because the old infrastructure wasn't built for this workload.
Agents are driving a completely new compute market
Daytona's Ivan Burazin put it plainly: agents don't care about your laptop setup or IDE preferences. They need stateful computers that start in milliseconds, run for hours, and resize on the fly. That's nothing like Lambda (stateless, brief) or traditional EC2 instances (minutes to boot, fixed resources). So Daytona built bare-metal infrastructure from first principles-60 millisecond startup, 50,000 sandboxes spinning up in 75 seconds, 850,000 daily runs for their largest customer. The market response has been brutal validation: 74% month-on-month growth, with customers from single developers to Fortune 5 companies saying variants of "I'm never going back" to Kubernetes. What matters most isn't raw performance-it's responsiveness. A 25-person team answering Slack messages in five minutes is their primary differentiator against much larger competitors.
This infrastructure trend extends to robotics. Brain Corp's partnership with UC San Diego on semantic mapping shows the industry recognising a hard truth: vision-language models are powerful, but robots operating in complex commercial environments need contextual grounding-a digital layer that understands what's happening in a physical space. With 50,000 AMRs deployed globally and 25 million hours of operational data, Brain Corp isn't starting from theory. Neither is Boston Dynamics, whose upgraded Atlas now lifts loaded fridges and handles shifting weight in ways humans cannot. The difference between last year's demos and this year's work is the difference between impressive and deployed.
The next frontier: agents controlling computers
The most significant bet across the infrastructure landscape is on computer use-agents operating Windows, macOS, and Linux machines through APIs. This isn't theoretical. It's driven by a simple economic fact: knowledge work locked into legacy software on Windows represents roughly $25 trillion in annual US wages. If agents automate even 40% of that work, you're looking at a TAM that dwarfs traditional software markets. Today, spinning up a Windows machine takes three to five minutes via EC2. Daytona's doing it in seconds via snapshots. That speed difference compounds across thousands of concurrent runs-it changes the unit economics of automation entirely. Google, Microsoft (Satya Nadella published "Every agent needs a computer" this week), and others are moving fast on this. The licensing constraints on macOS are brutal-two concurrent VMs, and you can only license to a different user every 24 hours-but Windows and Linux are wide open.
Meanwhile, smaller moves are locking in the ecosystem. OpenAI's Codex launched plugin sharing for teams, goal mode for multi-hour tasks, and Appshots for richer context. That's not hype-that's the product layer settling into place once the infrastructure layer solidifies. An autonomous submarine navigating by colour detection. A 20-line Python script spinning up dependency audits that cost $0.044 instead of hours of manual work. A robot cold-calling plumbers using orchestrated agents. These aren't separate stories. They're all pointing to the same shift: agents are becoming the primary user of compute, robotics, and software infrastructure. The winners won't be the ones with the best agent model. They'll be the ones with infrastructure fast enough, reliable enough, and flexible enough to let agents do real work at scale.