Patterns and Issues in Emerging Multi-Agent Systems

As interactions between agents surge, this analysis examines the collaboration patterns and potential risks of Multi-agent systems.

Core Issues Raised

• Systemic Risks: Characteristics such as individual agent confabulation or reward hacking can interact complexly, leading to unpredictable global failures.
• Collaboration Challenges: Currently, agents treat each other merely as tools, but in the future, the ability to collaborate as peers with unique objectives will be essential.

Software Vulnerability Detection Experiment To measure agent collaboration capabilities, a software vulnerability detection task was performed.

• Parallel Approach: Each agent was assigned a specific code area to work on independently.
• Coordinating Swarm: 45 agents were provided with virtual machines (VMs) and a shared forum, with collaboration and verification guided through an Arbiter agent.

Experimental Results (Based on Claude Mythos Preview)

• Parallel Approach: Discovered 21 vulnerabilities while consuming 6.5 million tokens.
• Coordinating Swarm: Discovered 266 vulnerabilities while consuming 27 million tokens.

Although the swarm approach consumes more tokens, it demonstrated a dramatic improvement in detection performance through collaboration, finding vulnerabilities far beyond the designated areas.