A harness that keeps task state outside the context, with a read-only auditor

This item is written from the paper's abstract. The 29-page body is unread by this edition. On that basis: the authors' complaint is that existing harnesses "maintain task execution, task state, and completion assessment within a growing context," which lets an incorrect self-assessment propagate into later decisions. Their answer is to hold task state outside execution and update it "only with facts independently verified from the environment." The Manage-Execute-Audit loop splits the work across three roles: a manager holds the state and picks the next subtask, a fresh-context executor performs it, and a read-only auditor checks the environment before the next round. Qwen 3.7-Plus goes from 51.8 to 80.7 percent on WeaveBench and 2.8 to 8.3 on OSWorld 2.0; Claude Opus 4.7 goes from 20.0 to 34.3 percent on an OSWorld 2.0 subset. These are the authors' own numbers on their own system, with no independent replication.

Why it matters: The pattern is separable from the paper and you can build it without the paper. If the agent that did the work is also the agent that decides the work is done, what carries forward is its self-assessment, not the result. A read-only auditor with a fresh context, checking the environment rather than the transcript, is the cheap version of that fix. Note the axis these numbers move along: same models, different harness.