LongHorizon-Harness pulls agent task state out of the context, lifting WeaveBench to 80.7%
Ziyu Ma, Hailang Huang and co-authors recast long-horizon agent failure as task-state management, holding state outside execution in a Manage-Execute-Audit loop: a manager picks the next subtask, a fresh-context executor runs it, and a read-only auditor verifies the environment before the next round. On Qwen 3.7-Plus it lifts WeaveBench from 51.8% to 80.7% and Terminal-Bench 2.1 from 69.7% to 77.2%, and a lightweight AgentAdapter swaps model and harness backends without modifying their native agent loops — portable to an existing agent stack. OSWorld 2.0 moves 2.8% to 8.3% and Claude Opus 4.7 goes 20.0% to 34.3% on a subset, so the gain is solid on terminal and tool work and still thin on full desktop control.
Source: arxiv.org ↗
maintains the task state explicitly outside execution and updates it only with facts independently verified from the environment
Why this matters
- → Separates task state from execution context, cutting long-horizon agent failures in half
- → Portable adapter swaps models and harnesses without native code changes
- → 80.7% on WeaveBench—a concrete baseline lift for agentic reasoning benchmarks