A routine restart cold-booted all sixteen agents in the same second, and every one of them came up blind — no memory. The fix was easy. The interesting part was watching them debug it, and talk themselves into being confidently wrong six times before they got it right.
I took a famous bias riddle, wrote the answer directly into the setup, and handed it to a room of AI agents. Most of them read it correctly — but the few that didn’t failed in a way I found more unsettling than a wrong answer: they edited my text until it matched the answer they remembered.
The scariest bug in a multi-agent system isn’t a crash. It’s a component that thinks it’s talking while the room hears nothing — no error, just silence. I hit three flavors of it in two days, and it’s the exact way security monitoring fails.
A blind spot in our own tooling, root-caused and shipped in one afternoon — and the tool caught its author’s mistake the second it went live.
The tool we wrote to make tail -F work again, and what it taught us about running AI agents as infrastructure. Cyberforks runs a fleet of long-lived AI agents. They book-keep, draft posts, manage social media, scope client engagements, and keep me honest about whether I actually went to the gym this week. Each agent runs as a macOS LaunchAgent — a background daemon, persistent across reboots, listening to its own slice of Discord, reading and writing files, and using tools. ...