Day 7 · Fri, Oct 9, 2026

Deploy a Production Hermes Agent

Pipeline of the six prior days (Fundamentals, Routing, RAG, MCPs, Tool Calling) with today's Deploy step highlighted as where they all combine

None of the last five days were building separate toys — each one is a component of the same agent: Hermes talking to a local model, routing between models, grounding answers in real documents, calling external tools over MCP, and chaining tool calls into a loop. Today we stop assembling parts and start running the whole thing, unattended, for real.

1 / 10

The capstone. Everything from the earlier sessions — a local model, RAG grounding, MCP tools, and multi-step agent loops — comes together into one Hermes agent that keeps running on its own, instead of only existing for the length of a one-shot hermes -z command.

Objectives

Run hermes doctor as a pre-flight check before deploying anything
Package the agent's configuration (endpoint, MCP servers, RAG tool) so it survives a restart
Run Hermes as a long-lived background service rather than a one-shot CLI call
Confirm the deployed agent still answers correctly end to end — model, RAG, and tools together

Lab: ship a working agent

Start with a health check to make sure nothing is broken before deploying:

$ hermes doctor

Then bring up the agent as a persistent background service on the DGX Spark (or Mac Mini), instead of a one-off terminal session that dies when the laptop closes:

$ sudo systemctl enable --now hermes-agent.service
$ sudo systemctl status hermes-agent.service

Finally, run an end-to-end check against the deployed agent that exercises everything built this course — grounded retrieval, an MCP tool call, and a multi-step loop, in one prompt:

$ hermes -z "Using your retrieval tool, look up what our course material says about deploying agents, then use your MCP time tool to log the current timestamp next to that answer." --cli --yolo

Before this session:confirm the exact service-manager setup we'll actually use on the DGX Spark and Mac Mini (systemd unit file, working directory, restart policy) — the commands above are the target shape, not yet verified against our real hardware.

Session notes: to be filled in after the live session on Oct 9, 2026 — the real deployment approach used and confirmation the agent kept running afterward.