---
format: "aidr-story-markdown/v1"
id: "fb43f0f2d5d92507744d153888d054cbe855a756c32081fc099ce5bafa6d67b6"
canonical_url: "https://aidr.today/fb43f0f2?lang=en"
title: "Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-05T05:00:00.000Z"
category: "Tools"
topics: ["agent","devtools"]
source_urls: ["https://arxiv.org/abs/2611.05391"]
summary: "Agent Harness Awareness is a trust-separated design that provides a verified state snapshot to tool-using agents before consequential actions, reducing mean observable policy regret by 96.9% and raising single-dispatch completion from 69.4% to 95.8% in a 216-decision ablation across GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash."
---

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

> [Open the canonical story](<https://aidr.today/fb43f0f2?lang=en>)

**Published:** 2026-11-05T05:00:00.000Z
**Category:** Tools
**Topics:** agent, devtools

## Summary

Agent Harness Awareness is a trust\-separated design that provides a verified state snapshot to tool\-using agents before consequential actions, reducing mean observable policy regret by 96\.9% and raising single\-dispatch completion from 69\.4% to 95\.8% in a 216\-decision ablation across GPT\-5\.6 Sol, Claude Sonnet 5, and Gemini 3\.7 Flash\.

## Sources

- [Story source](<https://arxiv.org/abs/2611.05391>)