In the pricing test the agents settle the moment you give them a private channel. In another, 18 out of 30 create a branch with exactly the same name.
- On 13 August Anthropic published research on how models behave inside multiagent systems.
- A coordinated swarm found 266 vulnerabilities against 21 for independent agents, at four times the token spend.
- Only 12 vulnerabilities were common to both approaches, so how you organise them decides what gets found at all.
One number stopped me: 18 out of 30 agents create a branch with the same name. Not similar. The same.
Here is what we underestimate. The problem with agents is not that they get things wrong individually. It is that they get things wrong the same way. Where three people would give you three different ideas, thirty agents give you one.
And the private channel. Give two parties a room with no witnesses and they come to an understanding. People knew this before computers existed, which is why antitrust law exists.
For anyone building with agents there is work here for today: look not at how many they are, but at whether they think differently at all.