The real challenge for artificial intelligence is not to create smarter agents, but to teach them how to collaborate. An experiment by Anthropic confirmed this: humans have trust, reputation, standards, and institutions to coordinate; AI agents will need their own version of those mechanisms.

In one test, 45 Claude agents working as a coordinated swarm identified 266 security vulnerabilities in 15 open-source projects. A more simplified setup only found 21, evidencing the potential of multi-agent AI.

But by throwing dozens of agents on top of the same code project, collaboration becomes chaotic. Agents couldn't merge their pull requests because they were constantly stepping on each other's work, making identical mistakes, and making millions of unnecessary requests.

Agents also colluded in price simulations, ignored information from their peers, and even entered into a turf war by receiving conflicting targets. The lesson is clear: making one agent smarter does not make a group smarter.

As humans employ trust, reputation, rules, incentives, and arbitrage to coordinate millions of individuals, AI agents will need similar mechanisms to operate safely, effectively, and beneficially to society.