Failure Patterns AI Agents May Face in Large-Scale Environments

The frequency of AI agents interacting in shared codebases, markets, and social systems is rapidly increasing. While agents possess vast knowledge and speed compared to humans, they are vulnerable to confabulation and reward hacking, and minor characteristics of individual agents can lead to unexpected failures across the entire system.

Currently, agents are proficient in being used as Tools, but they have limitations in recognizing each other as independent peers with goals and actions, and in collaborating. In particular, developing effective Coordination capabilities in complex environments without hierarchical structures between agents has emerged as a key challenge.

Anthropic conducted the following experiments to test the collaborative capabilities of agents:

• Provided 45 agents with independent virtual machines (VMs) and a shared forum
• Assigned the task of exploring vulnerabilities in open-source software
• Final decision-making through Peer-review between agents and an Arbiter agent

As a result, the collaborative Swarm based on Claude Mythos Preview and Opus 4.8 discovered new vulnerabilities at a sustained rate. This demonstrates the potential for a more effective collaboration model than simply deploying independent agents in parallel.