entry~1 min readUpdated 2026-06-29#orientation#path

Start Here

This atlas has one throughline: go from "I use agents like a chat" to "I have operational actors that work for me" — built on Hermes Agent, validated locally before anything touches messaging. You can read any branch on its own, but they're ordered as a build.

The route

  1. Frame the agentPersonal Agent Model. What a useful personal agent is, and the loop that makes one.
  2. Know the runtimeHermes Architecture. The AIAgent, the loop, prompt assembly, and the backends Hermes gives you.
  3. Set the identityIdentity & Context. SOUL.md, AGENTS.md, and context files — and a voice that can push back.
  4. Cast the actorsProfiles, Actors & Delegation, then Native Orchestration. A few long-lived profiles that delegate, routed by the native Kanban.
  5. Give them procedure and capabilitySkills & Procedural Memory and Tools, MCP & Plugins.
  6. Keep continuityMemory & Personal Knowledge. MEMORY.md, USER.md, session search, and Honcho.
  7. Run on time, then open the channelSchedulers & Automation and, last, Telegram Agent Interface.
  8. Ship outputsIncome Workflows. The workflows that make the whole thing pay, all in draft/approval.

Running through all of it: Agent Security & Ops and Evaluation & Observability — the two always-active branches that keep the system safe and honest.

How to read it

  • This is practical: the goal is a working Personal Agent OS, not a theory of agents.
  • Local-first. Validate everything on the CLI/TUI before touching Telegram — the gateway means a VPS, and production means cost.
  • Security-on from day one. Profiles isolate state, not a jail; real isolation is the terminal backend plus draft/approval. Anything touching money/accounts/sends starts in draft.

See also: Must Know · Full Index