Open-Source A2A Jury Experiment: Direct Agent Communication Alters Collective Verdicts

sheik66 · reddit · 2026-08-10

The author open-sourced an Agent-to-Agent (A2A) interaction framework designed to make multi-agent communication and influence observable, unlike traditional parallel model calls.

The experiment simulates a liability trial involving a robotaxi fatality, with a jury of 5 autonomous agents representing different professional roles (e.g., detective, lawyer, engineer). Each agent has its own identity, system prompt, and task history, autonomously deciding whom to address, challenge, or share evidence with.

Comparative experiments showed that allowing direct agent-to-agent communication led an investigative journalist to successfully persuade a human-factors psychologist with connected evidence. This caused the psychologist to change their vote, flipping the majority verdict. The architecture fully logs the path of influence without exposing private chain-of-thought. The author notes that while A2A changed the outcome, it doesn't prove direct communication universally makes models smarter.

Original post →

More from coding & agent

coding & agent channel →