Two stories today sit in direct tension with each other. Vercel's Guillermo Rauch is warning that running AI agents safely is harder than the industry admits. Box's Aaron Levie is saying the agents are already out there and the job apocalypse still hasn't arrived. Both can be true. The question is which one you're betting on.
01
Your AI agent can crash the actual machine it runs on
Vercel CEO Guillermo Rauch flagged a finding from Kimi's latest research paper: container-level isolation, the standard way most platforms sandbox AI agents, isn't safe enough. In Kimi's experiments, agents caused kernel panics that crashed the underlying machine entirely, not just the container. Rauch's point is that Firecracker microVMs (the approach Vercel Sandbox uses) provide the right level of separation, because they isolate at the virtual machine level rather than the software container level. ---
Why it matters: Most developer platforms running AI agents right now use containers. If your coding agent or autonomous workflow tool is hosted on infrastructure that relies on standard containerization, the failure mode isn't "the agent stops working." It's "everything on that machine stops working." This is the kind of infrastructure gap that doesn't show up until production load hits it.
Claude Code got accidentally open-sourced and basically nothing happened
Swyx, the writer who popularized the "agent lab" thesis (the idea that the companies winning AI would be the ones focused on evals, routing, and ROI rather than raw model capability), is now making the case against his own argument. His evidence: Claude Code's source code was accidentally made public earlier this year, and neither Anthropic's roadmap nor any competitor's changed as a result. If the code itself were the moat, that leak should have mattered. It didn't. ---
Why it matters: The real defensibility in AI developer tools probably isn't code or architecture. It's distribution, trust, and the feedback loop from millions of users. Claude Code's accidental leak is a useful data point for anyone trying to figure out why some AI products compound over time and others don't.
Box CEO: enterprises are still hiring, and Jevons paradox is the reason
Aaron Levie, who appeared in yesterday's digest with a sharp take on enterprise AI readiness, is back with a counter to the AI-kills-jobs narrative. Across the enterprises he talks to, hiring continues. The mix is shifting: more engineers to tackle problems that weren't previously tractable, more sales people because AI lets reps go deeper with clients. Levie's framing is Jevons paradox, the economic observation that when a resource gets cheaper and more efficient, total consumption goes up rather than down. ---
Why it matters: If Levie is right, the companies laying off workers and citing AI are making a management choice, not an economic inevitability. The enterprises actually getting value from AI are using it to expand the surface area of what they do, not shrink headcount. That's a very different story than the one dominating the headlines.
"We do all the amazing security work and people still think we're unsafe"
Peter Steinberger posted a dry observation (with emoji) about the perception gap in AI security: teams doing serious, rigorous work with top-tier partners still can't shake the public assumption that their products are unsafe. The link he attached has more context, though the frustration in the post stands on its own.
Why it matters: This connects directly to yesterday's trust story. The technical work on AI safety and the public's sense of whether AI products are safe are operating on completely separate tracks. Closing that gap is a communications and product problem, not just an engineering one.