ProtoLink团队在Hacker News上展示了一款新型AI多智能体（A2A）系统，将自主AI Agent置于虚拟法庭场景中，用于研究Agent间通信如何影响集体决策过程[1]。该系统的核心在于探索Agent之间的互动机制，而非具体案例本身[1]。
系统支持四种不同的通信拓扑模式进行对比分析：独立、直接网格、星型以及仅接收公开记录[1]。在确定性离线运行的默认模式下，系统生成JSON结果、ProtoLink追踪、公开记录和交互式HTML报告，使用户能够追踪整个决策过程[1]。一个具体案例显示，在网格拓扑中，法官Sofia Bell对心理学家Dr. Anika Rao的直接质询导致该角色的内疚度从77.90变为81.41，最终投票从"非有罪"改为"有罪"[1]。
系统兼容多个LLM后端，包括OpenAI、Anthropic、Gemini和Ollama等[1]。用户可通过--action-parse-attempts和--max-attempts参数控制验证流程，参数范围为1至5，默认值为3[1]。
The ProtoLink team has unveiled a replayable artificial intelligence multi-agent system that places autonomous AI agents in a virtual courtroom setting to study how inter-agent communication shapes collective decision-making [1]. The core innovation centers on agent-to-agent interactions rather than the specifics of individual cases, offering researchers a controlled environment to observe communication dynamics [1].
The system supports four distinct communication topologies for comparison: independent agents, direct mesh networks, star configurations, and scenarios where agents access only public records [1]. By default, the simulation runs deterministically in offline mode, generating comprehensive outputs including JSON results, ProtoLink traces, public records, and interactive HTML reports that allow users to replay and analyze the entire decision process [1]. The platform accommodates multiple large language model backends, including OpenAI, Anthropic, Gemini, and Ollama [1].
To demonstrate its capabilities, the team documented a notable outcome in the mesh topology configuration: when Judge Sofia Bell directly questioned psychologist Dr. Anika Rao, her guilt assessment shifted from 77.90 to 81.41, resulting in a voting change from "not guilty" to "guilty" [1]. Users can fine-tune agent behavior through parameters such as --action-parse-attempts and --max-attempts, which control validation attempts and accept values ranging from 1 to 5, with a default setting of 3 [1].