Your Agent Skills Should Outlive Your Vendor

On Thursday, November 12, Microsoft had a global outage. On Friday, November 13, my OpenAI keys were revoked for reasons outside my control.

My first reaction was oh no, there goes my productivity for the day. I use Codex to create tickets and to work across Confluence, Jira, GitHub, and Slack. It's how I make faster, better-informed decisions. Losing it felt like losing a limb.

Then I remembered I had the same agent skills configured in Cursor. I switched and kept working.

The lesson isn't that Cursor was a good backup. The skills were portable. The workflows I'd encoded (how I write a ticket, how I summarize a Confluence page, how I check a PR against its acceptance criteria) didn't belong to any one tool. They were instructions I owned, and any agent that could read them could run them.

Two kinds of failure

These two outages were different:

• The platform goes down. That's the Microsoft outage. Nothing you did caused it, and you can only wait it out or route around it.
• Your access disappears. That's the revoked keys. The service is fine, but you can't reach it.

A backup plan has to cover both, and most people only plan for the first.

What I'd tell anyone building on agents

1. Keep your skills in a tool-agnostic format that more than one agent can load.
2. Configure the fallback before you need it. I only avoided a lost day because I'd done this already.
3. Check what your backup depends on. If both agents call the same model provider, you have one point of failure with two logos.
4. Have a second backup. Mine worked, but I had exactly one. That was luck as much as planning.

When things move this fast, the speed isn't the asset. The asset is being able to keep moving when something breaks.