agent-verification

GoDaddy的代理名称服务（ANS）使用密码学，实现AI代理身份的离线、亚毫秒级验证，增强安全性并补充如OAuth 2.0等授权框架。

Presents a verification instrument for long-horizon agents that structurally separates commitment drift from binding drift, using a deterministic executive and pre-registered predictions. Reports ablation results showing commitment mechanism removal flips goal abandonment from 0 to 1 while binding error stays flat, though task efficacy is null on ARC-AGI-3.

讨论验证智能体在长时间任务后输出结果的困难，并提出是否使用批评者智能体或可追溯性工具来确保可信度。

文章描述了构建一个'harness'系统来为编码代理提供上下文、工具、来源和验证的经验教训，并详细介绍了八个支柱中的前两个：上下文和来源。