LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

2026-08-03

AI Summary

This work addresses the challenge that existing agents struggle with long-horizon tasks due to tight coupling between task states and execution contexts, which obscures state tracking and propagates erroneous self-assessments. The paper reframes this issue as a task state management problem and proposes explicitly maintaining an external task state updated solely based on environment-verified facts. To achieve this, the authors introduce a Manage-Execute-Audit (MEA) loop mechanism alongside a lightweight AgentAdapter architecture, enabling decoupled state representation, verifiable state updates, and plug-and-play compatibility between models and frameworks. Experimental results demonstrate substantial performance improvements across multiple large language models on benchmarks including WeaveBench, Terminal-Bench 2.1, and OSWorld 2.0, with gains of up to nearly threefold in some cases.

Abstract

Large language model (LLM) agents increasingly undertake long-horizon tasks that require sustained reasoning, tool use, and revision across many interdependent steps. However, existing agent harnesses maintain task execution, task state, and completion assessment within a growing context, making the state difficult to track and allowing incorrect self-assessments to propagate into later decisions. We reformulate long-horizon execution as a task-state management problem and propose LongHorizon-Harness, which maintains the task state explicitly outside execution and updates it only with facts independently verified from the environment. Its Manage-Execute-Audit(MEA) loop uses a manager to maintain the task state and determine the next subtask, a fresh-context executor to perform it, and a read-only auditor to verify the resulting environment state before the next round. A lightweight AgentAdapter supports interchangeable model and harness backends without modifying their native agent loops. LongHorizon-Harness improves Qwen~3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench~2.1, and from 2.8% to 8.3% on OSWorld~2.0. It also raises Claude Opus~4.7 from 20.0% to 34.3% on an OSWorld2.0 subset, demonstrating consistent gains across models, harnesses, and interaction domains.

Problem

• long-horizon tasks
• task state management
• self-assessment propagation
• context growth
• agent execution

Innovation

• Long-horizon agents
• Task-state management
• MEA loop
• Environment verification
• AgentAdapter