Tag
该论文提出'standing'概念，研究多用户聊天中智能体的承诺能否被非债权人取消，并在六个模型上区分未授权释放与覆盖两类错误，发现自维护承诺账本加释放策略可显著提升严格准确率。
This paper introduces TRACE (Typed Reasoning And Commitment Evidence), a typed, versioned schema for recording reasoning traces in agentic systems to enable auditability and improve reasoning quality. It defines a reference writer, measurement regime, and consumer contract, with two worked examples illustrating the approach.
A discussion on how teams handle AI agents making real commitments without human approval, seeking exceptions and insights on liability and legal friction.