OpenClaw + Claude Code

Level 1 · Course 6

One agent is a single point of failure. When it's stuck, you're stuck. When it's frozen, you can't reach the controls. When it builds something wrong, it's the same eyes reviewing the same output. You need a second agent that doesn't depend on the first — for builds, for QA, for reaching the config when the gateway's dead.

This course covers the in-house architecture for a two-agent system — builder and reviewer running on the same machine, separate auth, clear boundaries, one handshake to authorize everything.

What this course installs: Claude Code CLI — authenticated, verified, running on your machine. Shepherd Protocol — full-authorization handoff for gateway restarts, config changes, and production deploys. Two-agent boundaries — who builds, who manages, who fixes what. Crash protection rules — config changes route through Claude Code before they're applied. Survival test — verify Claude Code is alive and reachable after every session start.

When Two Agents Matter

Two agents matter when your main agent is the single point of failure for config changes, production deploys, or anything where a mistake means downtime. A builder agent with separate auth can restart the gateway, edit config files, and deploy without depending on the main agent being functional. If the main agent is down, the builder is still reachable.

Two agents are overkill when you're doing research, writing, planning, or anything where a mistake is reversible in 30 seconds. One agent is fine. Save the second agent for the work where something breaking actually costs you.

The Scenario

An operator's agent suggests a config change. Without a second agent, the main agent edits the config, restarts the gateway, and if something breaks — the main agent can't fix it because the gateway it depends on is down. The operator is locked out. With a builder agent, the main agent hands off the edit. The builder makes the change, restarts the gateway, and verifies the result. The main agent stays up the entire time. One point of failure becomes two independent paths.

You Decide

Is your agent currently your only path to config changes? If yes, you need a second agent. If you can reach config through a terminal or a different device, you already have a backup — the second agent becomes optional. Decide which risk level you're at before you install.