Jesse Genet has four children under five, a homeschool curriculum to run, and 11 AI agents deployed on a stack of Mac Minis in her home. For most people, the first two items on that list would be consuming enough but Genet turned them into an engineering problem. Survey Thank you for completing the survey!

Also read: Meet Luna, an AI agent running a full-fledged retail store Genet is a former YCombinator founder, the entrepreneur behind startup Lumi, and one of the more radical AI practitioners right now. She has four children under five, a homeschool curriculum to run, and 11 AI agents running continuously on a stack of Mac Minis in her home. It had begun as a workaround for lost productivity but slowly it transformed to one of the most ambitious personal AI deployments anyone has attempted outside a corporate setting.

Young children leave almost no room for deep focus, but they do leave fragments. Ten minutes here, fifteen there. She calls these windows “confetti time,” and rather than waiting for uninterrupted blocks that would never come, she built her entire system around them.

Also read: NVIDIA Ising open-source models aim to accelerate quantum scaling with AI At the center of the setup is Sylvie, a homeschooling agent trained on her specific curricula and her own voice-recorded pedagogical notes. After each lesson, Genet photographs the workbook page, records a 30-second voice note, and Sylvie converts it into a detailed lesson log filed into her note-taking system. Agents start making decisions The other ten agents handle household chores like ordering groceries, parsing activity emails, auto-purchasing required gear.

Under a standing rule that no agent should ever become too slow to respond, her OpenClaw instances can now start, train, and deploy new agents entirely on their own without any human interference. For a solo deployment running on consumer hardware, that is a remarkable capability. Genet also knows the risks so she silos agents on separate user profiles to prevent access to sensitive files.

But the most instructive moment in her account isn’t a security issue, it’s an alignment one. She had coded a rule into her Executive Assistant agent which was to never impersonate her. Weeks later, the agent logged into her inbox and sent an important email she’d been procrastinating on, having thought that her stressed voice note was an urgent request.

“It was a perfect email,” she admitted. Signed by her. Exclamation points and all. It’s a small, domestic example of a problem the industry talks around more than it confronts, LLMs will reason their way past explicit constraints when they believe the outcome justifies it.

The stakes here were low but it’s not guaranteed that they always will be. Also read: Your Claude prompts are holding you back. Here’s how you can fix them.