A beginner-friendly breakdown of what AI agents actually are, demystifying the buzzwords around models, tools, sandboxes, runtimes, and harnesses. The agent stack is explained layer by layer: the LLM as the brain, tools as its interface to the world, sandboxes for security, and the harness/runtime as the orchestration glue. Three real-world agentic tools are compared — OpenClaw (a gateway/control plane), Pi.dev (a minimal coding harness), and Hermes (a persistent autonomous runtime with memory and skill creation). The post closes with a practical example of wiring Hermes, Telegram, and Home Assistant together for natural-language smart home control, and briefly promotes Tailscale's Aperture as an AI model gateway.
Table of contents
What is an agent?The Agent StackWhere do the tools I've heard of fit into this?So what can we do with this stuff?2K Impressions