What this is

This is a record of how one engineer runs a 24/7 multi-agent AI stack on bare metal: long-lived agents (OpenClaw, Hermes) working in tandem with daily coding harnesses (Codex CLI, Claude Code, OpenCode). It is not a framework and not a tutorial series. Every recipe is deployed in production on real hardware, and the failures are documented next to the fixes.

The core idea: keep agents aware of work happening across every harness by maintaining synchronized durable context instead of isolated session memories. Memory files, tool catalogs, and agent instructions are treated as operational infrastructure, with the same hygiene as backups and secrets.

Who it's for

Engineers who run agents as long-lived services rather than chat sessions, who prefer owning their stack over renting it, and who want the post-mortems as much as the recipes. Everything here assumes a single operator, a Linux host, and a healthy distrust of letting any model touch production unsupervised.

License

The written guides are licensed CC BY-NC-ND 4.0. The drop-in templates are MIT. If you want to suggest a fix or a new recipe, see CONTRIBUTING on GitHub.

Related

• Brigade - the companion CLI for agent memory, handoffs, and guardrails.
• content-guard - the publish-boundary scanner used by this repo's pre-push hook.
• OpenClaw - the long-lived agent gateway most of these recipes orbit.