Paper 2608.01964
LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
- Published
- Aug 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 311 stars
01 In brief
Summary
LongHorizon-Harness, from Alibaba's DreamX Team, addresses long-horizon LLM agent failures by reformulating execution as task-state management.
It uses a Manage-Execute-Audit (MEA) loop: a manager maintains explicit task state and defines subtasks, a fresh-context executor performs each subtask, and a read-only auditor independently verifies environment changes.
Only audited facts persist across rounds, preventing context rot and error propagation.
The framework improves Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, 69.7% to 77.2% on Terminal-Bench 2.1, and 2.8% to 8.3% on OSWorld 2.0, with gains also for Claude Opus 4.7.
Token costs vary by task and model capability, with the auditor being the main overhead.
The approach is most effective when tasks require preserving and revising multiple dependent environment states, while gains are limited when performance depends on a single model capability.
Agent capability is shown to be a system property of both model and harness, with a stronger harness raising the performance floor of weaker models.
02 From the paper
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.