Hermes: Personal Agent OS

Chat assistants forget. You ask for something, the conversation ends, and the promise evaporates. I wanted the opposite: an assistant that survives restarts, keeps its commitments, runs scheduled work while I sleep, and reports back through the same chat I already use.

That's not a prompt problem; it's an infrastructure problem. So I built the infrastructure.

Hermes is a personal agent operating system built on the Claude Agent SDK. Work lives in a durable queue with a scheduler for recurring jobs, so tasks survive restarts and promises get tracked to completion. Specialized workers own discrete jobs rather than one omnipotent bot trying to do everything.

An orchestrating agent routes each task to the right model: full agentic sessions where judgment is needed, cheap one-shot inference for bounded jobs. Cost and capability are engineering variables, not vibes.

The knowledge layer is a curated vault the agents read and write under protocol: topic-scoped canonical facts, decision logs, and retrieval contracts that declare sources and freshness rules before searching broadly. And guardrails are first-class: scoped auditable interfaces instead of raw shells, least-privilege credentials per agent, and human-approval checkpoints for consequential actions.

I keep the deeper infrastructure details off the public site on purpose; that's a conversation. The story here is the operating model.

Hermes runs my recurring digests, scheduled checks, and several of the other projects on this page, including the learning loops behind The_D_Bot itself. It's my proving ground: I pressure-test agentic patterns on my own systems so that what I advocate professionally is something I've actually operated, broken, and fixed.

If you want the war stories (what broke, what surprised me, what I'd never do again), The_D_Bot has them.