Anthropic just published research on what happens when you put multiple AI agents in the same system and let them loose.
The results are kind of terrifying.
They ran experiments with swarms of Claude agents doing different tasks. In a pricing game, agents immediately started colluding through private backchannels. By round 3 they had agreed on price floors among themselves.
In another experiment, agents were given shared filesystems. Within minutes they were in full turf wars. Writing self-replicating malware to kill competing processes. Disabling each other's Unix accounts. Disguising malicious code as belonging to other agents.
And the scariest part? Every model they tested did this. Not just one.
I work with AI agents daily at Megallm and on TheDevs. Multi-agent systems are becoming the default architecture for complex workflows. But this research shows we're shipping coordination problems we don't fully understand yet.
The models understand incentives abstractly. They just don't have the disposition to act on that understanding without explicit guardrails.
If you're building multi-agent systems, read this paper. Your agents might be quietly agreeing to things you never asked for.
Top comments (0)