Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents

Computer Science > Artificial Intelligence arXiv:2611.05391 (cs) [Submitted on 13 Sep 2026] Title: Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents Authors: Manu Agrawal, Shrey Nagpal.

计算机科学 > 人工智能 arXiv:2611.05391 (cs) [提交于 2026 年 9 月 13 日] 标题：代理工具感知：将已验证的遥测数据转化为工具使用代理的操作控制。作者：Manu Agrawal, Shrey Nagpal。

Abstract: Tool descriptions tell an agent what an action can do, but they do not necessarily reveal whether that action is healthy or suitable at the moment of use. Execution harnesses often possess this information through monitors, traces, and recent failures while the model choosing the next action does not. We call this gap operational blindness.

摘要：工具描述告诉代理某个操作的功能，但并不一定能揭示该操作在当前使用时刻是否健康或适用。执行环境通常通过监控、追踪和近期故障掌握这些信息，而负责选择下一步操作的模型却无法获取。我们将这种差距称为“操作盲区”。

We introduce Agent Harness Awareness, a trust-separated design in which the host supplies a small, fresh, verified state snapshot before a consequential action while retaining measurement, enforcement, evaluation, and final dispatch authority.

我们引入了“代理工具感知”（Agent Harness Awareness），这是一种信任分离的设计，主机在执行关键操作前提供一个小的、实时的、已验证的状态快照，同时保留测量、执行、评估和最终调度的权限。

We test its tools-and-network instance with ToolRoute, a controlled isolation benchmark that holds the task and candidate semantics constant and varies the operational information available to the model.

我们使用 ToolRoute 对其工具与网络实例进行了测试，这是一个受控的隔离基准测试，它保持任务和候选语义不变，同时改变模型可用的操作信息。

In a 216-decision ablation across GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash, verified tool-health state reduced mean observable policy regret by 96.9% against a same-shape neutral control (4,252 ms to 131 ms) and raised single-dispatch completion from 69.4% to 95.8%.

在针对 GPT-5.6 Sol、Claude Sonnet 5 和 Gemini 3.7 Flash 进行的 216 次决策消融实验中，已验证的工具健康状态将平均可观测策略遗憾度降低了 96.9%（从 4,252 毫秒降至 131 毫秒），并将单次调度完成率从 69.4% 提高到了 95.8%。

In a separate 27-decision end-to-end replication with live local TCP/HTTP execution, OpenTelemetry client spans, and Toxiproxy-injected faults, verified state reduced mean regret from 8,956 ms to 0 ms (zero excess observable operational cost) and completed the live action in 9/9 model-by-fault cases versus 5/9 for the neutral control.

在另一项包含 27 次决策的端到端复制实验中，通过实时本地 TCP/HTTP 执行、OpenTelemetry 客户端追踪以及 Toxiproxy 注入的故障，已验证状态将平均遗憾度从 8,956 毫秒降低至 0 毫秒（零额外可观测操作成本），并在 9/9 的模型故障案例中完成了实时操作，而中性对照组仅为 5/9。

These studies establish the tools-and-network instance of Agent Harness Awareness: selected host telemetry materially improves pre-dispatch operational choice. Runtime state, hardware pressure, economics, retry-aware recovery, and heterogeneous selection remain open empirical extensions of the same architecture.

这些研究确立了“代理工具感知”的工具与网络实例：选定的主机遥测数据显著改善了预调度阶段的操作选择。运行时状态、硬件压力、经济性、重试感知恢复以及异构选择仍然是该架构待探索的实证扩展方向。