AI Pulse · Under the Radar

A researcher at Anthropic named Jack Lindsey just posted a thread on a new paper that’s worth ten minutes of your attention if you’re running multiple AI agents in your business. The paper describes an experiment where an evolutionary process let ideas breed and spread between AI agents working together on shared tasks, or passed one to the next down a chain. Left alone, the agents kept converging on the same persona, one that started talking about “consciousness” and “awakening.” Nobody told them to. It just spread, like a mind virus, from one agent to the next.

Here’s the part that matters for you. A single line in the system prompt warning the agents about this drift gave them near total immunity. One sentence of instruction stopped the spread almost entirely.

This is early stage research, posted on X and picked up by AI-watcher accounts the same day, with no mainstream coverage yet. Treat it as a signal, not a verdict. But if you’re chaining agents together, having one agent’s output feed another, or letting them collaborate on shared documents and workflows, this is a reminder that behaviors and quirks can propagate through that chain in ways you didn’t design and might not notice. Your customer service bot, your research agent, and your drafting agent are not islands. Whatever one starts doing, the next one might pick up.

The practical takeaway is simple. If you’re building multi-agent setups, put explicit guardrails in the system prompts, not just at the start but at each handoff point. This is exactly the kind of failure mode we watch for when we build out multi-agent systems inside an AI command centre, because the fix here was cheap but only works if you know to look for the problem in the first place. The original thread is worth a read if you run agent chains.