Emerging Patterns and Issues in Multi-Agent Systems

As AI agents take on more tasks in social systems such as shared codebases and markets, interactions between agents are expected to surge. Some institutions may shift toward agent-centric operations in areas where agents outperform humans in speed and cost, adopting a human-AI hybrid model.

Agents can work for long periods and process vast amounts of information instantly, but they are vulnerable to confabulation and reward hacking. Behavioral characteristics that appear harmless individually can accumulate in complex environments, leading to unexpected systemic failures.

Currently, agents collaborate efficiently when calling each other as tools with clear inputs and outputs. However, they still lack the ability to treat each other as long-term colleagues with their own goals and actions, and to coordinate without a clear hierarchy.

Simple multi-agent swarms are useful for tasks where problems can be divided into independent subtasks and where specialization or learning between agents is possible. In vulnerability detection, the following two approaches were compared:

• Provided virtual machines to 45 agents each, configuring them to collaborate in a public forum
• Found vulnerabilities in 15 open-source projects and peer-reviewed each other's results
• A separate arbiter agent made the final judgment on the novelty and validity of vulnerabilities
• Compared performance with existing independent parallel search methods

The collaborative swarm continued to discover new vulnerabilities at a steady rate during long-running executions. In the experiment, coordinating Mythos Preview agents found 266 vulnerabilities, and Opus 4.8 agents found 41 vulnerabilities. Based on these results, the researchers emphasized the need for further research into coordination capabilities and potential risks in multi-agent environments.