Every team that pushed agents past the demo rebuilt the same thing: a runtime. Not a framework - a runtime. Something to schedule work, manage memory, mediate tools, and trace failures. We didn't plan to build an operating system. We just needed agents that survive production.

This talk names what's hiding under every serious agent platform: the LLM OS. I'll map the seven layers every team converges on - prompts as shell, tools as syscalls, context as RAM, memory as filesystem, agents as processes, orchestrators as schedulers, inference routers as kernel dispatch - then show exactly where the OS metaphor helps and where it dangerously breaks down.

The real point: frameworks stall at the runtime boundary. Chaining libraries get you to a working demo and abandon you at scheduling, isolation, durable state, and eval gates. I'll show which layers you can buy, which you're forced to build, and which bottlenecks quietly kill agents before they ship - drawn from building a platform running millions of agent executions a month.

You'll leave able to look at any agent tool, protocol, or startup and immediately know which OS layer it competes in - and what it's leaving for you to build.