Learning Hermes Agent

I've been exploring Hermes Agent as a coding/operations partner, and these are the notes I wish I'd had on day one.

What Hermes Agent actually is

Hermes Agent is an autonomous AI assistant you run locally or on a remote VM. It pairs with you inside a codebase, runs terminal commands, edits files, and can delegate to sub-agents for long-running work. Think of it as a careful senior engineer that happens to be a model.

Skills are reusable procedures

The most useful concept is the skill. A skill is a written-down procedure for a recurring task type — deploy a Next.js app, review a PR, query an S3 bucket, generate an image, and so on. Once saved, the agent loads the skill's SKILL.md and follows its exact steps instead of improvising.

Rule of thumb: if you repeated a task more than twice, save a skill.

The wiki keeps knowledge connected

There's an interlinked wiki (the Karpathy-style LLM wiki idea) where notes link to each other. Writing a note is cheap; the value is in the links — you can traverse from "deploy Next.js" to "Hostinger object storage" to "presign S3 URLs" without re-deriving anything.

A few things I learned the hard way

• Verify, don't trust logs. A background process reporting "Ready" is not proof it's serving traffic — curl the endpoint.
• Externalize native modules. Packages like better-sqlite3 must not be bundled; let the runtime require them.
• Secrets belong in env, never in code. Use a kubectl create secret or a Docker env block, and keep .env.example tracked.

Try it

1. Stand up the agent on your VM.
2. Save your first skill for a task you do weekly.
3. Start a wiki note and link it to two others.

That's the loop: do the work, capture the procedure, connect the knowledge.