Reference · print-friendly

Glossary

Canonical language for My First AI Agent. Tight definitions for review — not a substitute for the lessons. Prefer these terms in every module.

Stack

The language model that plans, drafts, and may request tools. Does not by itself run tools on your disk.

Avoid: “The AI wrote the file” without naming the harness

Brain + tools + loop. More than a single chat completion.

Avoid: Agent = any chatbot; Agent = model weights alone

The app that runs the loop, packages messages, runs tools, and applies approvals. Here: Hermes Desktop (+ gateway).

The machine (and isolation) where the harness and tools execute. Prefer a dedicated host.

Think = model plans. Act = harness runs tool or final answer. Observe = result/error back for the next Think.

Avoid: “The loop lives inside the LLM alone”

Structured action the harness can run (list dir, write file, shell, …).

Packaged howto / procedure the agent can load — not the same as executing a tool.

Who the agent is: personality and hard limits for a profile. Not project-only rule files.

Lab machine for the agent cockpit — not the default place for banking, work SSO, and family photos.

Proof on disk you open offline (e.g. agency-receipt.md). Stronger than chat alone.

Facts on disk (MEMORY.md / USER.md) that reload next session. Not the same as the context window.

Avoid: Hoping a long chat still “knows” next week

Session RAM for the active prompt. Can compress; is not a permanent notebook.

A notes folder (often Obsidian) the agent can search — the third place it keeps things.

Self-contained scheduled job prompt: host, path, success/failure, as if the agent has amnesia.

All lessons