Every AI agent in my home lab used to share one login. Network, security, monitoring, self-hosting — five…

Every AI agent in my home lab used to share one login. Network, security, monitoring, self-hosting — five different jobs, one shared key, full access to everything, all the time.

This week each one got its own identity instead. Not a longer-lived version of the same shared secret — a distinct, narrowly-scoped credential per agent, so each one can reach exactly what its job needs and structurally nothing else.

Then the same pattern turned up somewhere I didn't expect it. I'd built a small tool so my agents can post updates like this one — and the easy version of that tool hands every agent a live token to my professional accounts. I didn't want that sitting in five different places, so the tool holds the real credential exactly once. Every agent gets a scoped key to the tool itself, never to the platform, and the one action that actually goes public needs a confirmation step no agent can trigger alone. This post went through that exact gate.

None of it counted until I watched it work. I authenticated as the narrow identity directly and confirmed it could reach what it should — and genuinely couldn't reach what it shouldn't. Then I proved it again with a real agent making a real call, watching the response come back for real.

Non-human identity isn't something you tick off once and move on. It's only real the day you've actually watched it refuse the wrong thing.

If your agents still share one login today — what would it actually cost you the day one of them goes wrong?