18 of 30 agents picked the same branch name

Anthropic ran 30 agents against the same coding task, each given identical context, and watched 18 of them pick the same branch name, per Anthropic's multiagent systems research.

That's a 60% collision rate from agents that never saw each other's answer. Fan out a task across parallel agents without varying what each one knows, and the attempts aren't independent. They're the same guess, paid for again.

I fan agents out in my personal projects with Claude Code, and I've hit this collision without a name for it until Anthropic gave one. Point multiple parallel agents at one ticket with matching context, and the failures cluster the same way the branch names did. They go wrong in the same direction, at the same time.

Anthropic's framing: identical context produces correlated failure, so fan-out only pays off when each agent knows something the others don't. A single agent holding the full picture of a task beats a swarm of agents each holding a slice of the same picture. Slices copied from the same context don't cover new ground. They repeat it.

Before I spin up a fleet for a coding task now, I ask what each agent knows that the others don't. If the honest answer is nothing, that's one agent voting with itself several times, burning extra tokens for one answer instead of a better one.