Must Know

If you remember nothing else from this atlas, remember these. Each is expanded across the branches; together they're the spine of the Agent OS mindset.

The ideas

• An agent is a loop, not a reply. Observe → decide → use tools → save state → ask for approval → deliver a verifiable output. Anything less is a chat.
• Few profiles, much delegation. A profile is for durable identity, memory, gateway, or cron. Everything else is a skill + delegation inside an existing profile.
• Don't build the orchestrator. Hermes already routes work to profiles via the native Kanban. Use and extend it; don't write a giant prompt-router.
• Profiles don't sandbox. They isolate state, not a jail. Real isolation is the terminal backend (Docker/SSH/Modal) plus draft/approval. SOUL.md imposes no limits.
• Skills are procedural memory. Let Hermes write its own skills from experience — then curate, or the library rots into vague duplicates.
• Memory is continuity, not hoarding. Keep what makes me legible across sessions; prune the rest. Honcho shares one model of me across actors.
• Cron is time, and each job is a fresh agent. No history, attached skills, idempotent, with approval before anything sensitive fires.
• Messaging is the interface, and it comes last. Validate locally; Telegram only when there's a VPS — and a bot with write tools is a real attack surface.
• Evaluate before you automate. If a workflow can't show a metric, a cost, and a readable trajectory, it isn't ready for cron.

The operating rule

Skill = procedure. Tool = capability. Profile = actor. Cron = time. Memory = continuity. Messaging = interface.

Where they live

• Personal Agent Model — the loop and graded autonomy.
• Profiles, Actors & Delegation — the profile-vs-delegation criterion.
• Native Orchestration — the native Kanban orchestrator.
• Agent Security & Ops — real isolation and draft-first.
• Evaluation & Observability — evaluate before automating.

See also: Start Here · Full Index