Claude Went Down. I Opened Codex. Zero Downtime.

The setup

Yesterday I posted about rebuilding my workspace to be agent-agnostic. Plain markdown, plain YAML, env-var paths, no Claude-specific lock-in. The thesis: when the tooling layer churns, the workspace outlasts it.

I did not expect to test that thesis the next day.

What happened

Claude went down mid-task.

I opened Codex CLI in the same workspace. Same skills loaded. Same memory. Same briefs. Same manifests. Codex read the workspace exactly the way Claude reads it, because the workspace is just files.

The task done. Faster, actually.

Zero downtime. Zero re-plumbing. Zero "let me port my setup."

Why it worked

Three properties carried the swap.

The orchestration layer is plain text. Briefs, manifests, memory, voice rules. All markdown and YAML. Any agent that reads files reads my workspace.

The agent-specific bits are isolated. Hooks, slash commands, settings live behind one entry point. Codex doesn't need them. The rest of the system functions without them.

The skills are portable. My skill definitions aren't Claude-shaped, they're task-shaped. Codex picked them up and ran.

The lesson

When you build your stack around one tool, an outage is a stop. When you build your stack around your workspace, an outage is a tool swap.

The agent is a worker. The workspace is the contract. Workers are interchangeable.

This is the whole point of decoupling.

What I'd do differently

Nothing.

Yesterday's migration was the work.

Today's outage was the dividend.

If you're still running everything inside one agent's surface, that's a single point of failure dressed up as convenience. Pull your config, briefs, and memory into plain files. Put the agent-specific layer in a sidecar.

Test the swap before you need it.

You will need it.

// A<3

17 comments