From LLM Apps to Agent Systems: Building an AI Operating System with Octos

What if you could design agents like Socrates' daimonion — not one, but many? A beacon that watches your project evolve and generates a slide deck before Monday's meeting. Another that curates your morning news into a personalized radio station, read aloud in your preferred voice. A third orchestrating a research pipeline overnight with eight workers fanning out in parallel. Each arm thinking independently, all sharing one brain.

Octos — the Open Cognitive Tasks Orchestration System — makes this architecture real. A single 31MB Rust binary that serves 200+ users on a Mac Mini. No Python, no Docker, no runtime dependencies. 15 LLM providers with 3-layer failover so your daimons never go silent. 12 messaging channels so they meet people where they already are. Three memory layers so they remember what matters. And smart tool management that keeps agents sharp — 30+ capabilities available, but only 15 active at any moment, like an octopus choosing which arms to extend.

In this talk, we'll explore the design principles behind building agent applications that transcend chat — from multi-channel presence to proactive orchestration, from provider resilience to the hard problem of giving an agent many arms without losing its mind.

You'll leave with a new mental model: agents not as tools you use, but as daimons you design.