⚡ Activation

💭 in your head · nothing to type

Picture a theatre director: they hold the whole production, while actors receive only their scenes.

Which scene would you refuse to hand to one actor whole?

👁 Reflection

💭 in your head · nothing to type

Picture a theatre director 🎭 at rehearsal. They don't play every part in turn. They hold the whole production in their head and give each actor exactly enough to play their scene. The actors don't need to know the ending — the director does.

Return to the director: Hermes holds plan and state, while a worker receives one atomic scene.

Where in your task will you draw the boundary for the first scene?

💡 Concept

Director and worker

Context saturation appears when one agent tries to hold the whole implementation. The boss-worker model separates roles: Hermes as boss frames a subtask, launches a worker, receives and checks the result, then updates the plan. The worker context stays atomic and scoped.

How the author's setup works: an owner-controlled stack on personal hardware

Here Sovereign means your own stack: models on your hardware, data stays away from providers, and the fallback chain remains under your control. This is not a product description; it is the general design principle.

Author's programs: Claude Code, Codex, and Hermes
                         │
                         ▼
One model gateway: LiteLLM on an always-on Mac mini
  • programs use one endpoint through pool aliases
    (a pool = interchangeable models under one name)
  • provider keys live only on this machine
  • free provider tiers first; the local step comes last
                         │
             ┌───────────┴───────────┐
             ▼                       ▼
Ordinary tasks: cloud pools     Personal and medical data:
with fallback choices           a separate pool with no cloud neighbors
             │                       │
             └───────────┬───────────┘
                         ▼
Local model on a 128 GB Mac Studio
(an open model running locally)

The separate Linux machine is the console: gateway and monitoring panel.
Hermes lives on a separate machine running Linux.

Real route example: Claude Code → LiteLLM → Cerebras gpt-oss-120b. Hermes → LiteLLM with a fallback chain follows the same principle: Hermes dispatches work to executors (Claude Code, Codex), while the shared model gateway selects models.

What to install in practice

The official Hermes repository is NousResearch/hermes-agent. The official LiteLLM repository is BerriAI/litellm. Install Hermes and configure the model gateway by following their READMEs; the course has no public repository of its own, so this unit does not provide invented install commands.

A complete stack like this takes weeks, not an evening: several machines, a model gateway, model pools, monitoring, fallback routes, and rules for sensitive data. The course gives you the minimum experience: local LiteLLM on localhost:4000 (u3) + one local model through llama-server as the final step + Hermes on top, following the official instructions.