Tools and scripts to build images that turn ARM... AGENTS.md

This AGENTS.md turns an OpenClaw agent into a resident sysadmin for a single Ethereum node running on an ARM64 board. It is a good example of a rarely-seen pattern: an agent file written for an appliance rather than for a codebase. The agent has one machine to look after, root on it, and a standing job — keep the execution and consensus clients synced and alive — so the file is built around what to check and which command reveals it, not around coding conventions.

The most instructive move is the opening section. Before anything else, the agent is told to read two sibling files: the persona and safety rules, and the periodic health checklist. AGENTS.md stays short by delegating — persona lives in SOUL.md, the recurring checks live in HEARTBEAT.md, and per-client trivia (ports, flags, log patterns) lives in a references/ directory the agent reads on demand rather than carrying in its context.

How the file is organised

• Everyday Session — the read-these-first ritual that loads persona and checklist at the start of every run.
• Core Capabilities — the five things this agent is allowed to do: manage services, monitor health, read logs, self-heal common faults, and query the node's own APIs.
• Tools & System Info — a one-line-per-tool table of the exact commands that answer each question, from systemctl and journalctl through to reading the board's thermal zone.
• Monitoring & Metrics — where the Grafana dashboards and Prometheus alert rules live.
• Key Paths — a four-row table of the directories that matter: config, chain data, binaries, and the JWT auth secret.

What is worth borrowing

The capability list is written as an explicit allowlist, which is what makes the safety rules in the companion SOUL.md enforceable — anything not named is out of scope. The tools section pairs each utility with a concrete invocation (journalctl -u lighthouse-beacon-mev -n 100 -f) instead of describing it, so the agent has no reason to improvise flags. And routing hardware specifics to references/ keeps the root file readable as the number of supported clients grows. If you are writing an AGENTS.md for any long-lived machine — a home server, a build box, a NAS — this structure transfers almost unchanged.