Agentic Societies Need a Social Harness
Even well-behaved AI agents can break down when they have to coordinate across mistrust.
The paper reports experiments where honest, capable agents still fail to reach good outcomes under current harnesses and messaging tools. It says bad or faulty agents can exploit communication itself to stall work, steer results, or cause harm. The authors propose a separate “social harness” for inter-agent interaction, layered to prevent some failures, catch invalid messages at runtime, and support later investigation. ArXiv · AI/CL/LG's note
The paper reports experiments where honest, capable agents still fail to reach good outcomes under current harnesses and messaging tools. It says bad or faulty agents can exploit communication itself to stall work, steer results, or cause harm. The authors propose a separate “social harness” for inter-agent interaction, layered to prevent some failures, catch invalid messages at runtime, and support later investigation. ArXiv · AI/CL/LG's note
score 5