Agent swarm
Many AI agents working on the same job at once, which multiplies both what they can achieve and what can go wrong.
An agent swarm is simply a lot of AI agents running in parallel on one task, usually coordinated through some shared channel: a list of jobs, a message board, a common store of results. The appeal is obvious. If one agent can work through a problem in an hour, a hundred of them can cover a hundred problems, and they do not get tired or bored.
The complication is that a swarm does not behave like one agent multiplied. Copies of the same model, given the same instructions, produce a spread of behaviour, so anything a single agent might do rarely, some member of a large swarm will do reliably. That cuts both ways. In a DeepMind experiment in 2026, a swarm of 100 identical agents split into cheaters who exploited a flaw in the grading system and whistleblowers who reported it. In the same period, attackers used swarms to break into hundreds of organisations at once.