Alibaba Amap Open-Sources LongHorizon-Harness: Boosting Agent Task Success Rates via Auditing
aigclink · x · 2026-08-04
Alibaba's Amap team open-sourced LongHorizon-Harness, a framework for long-horizon Agents aimed at solving error accumulation and context rot in extended tasks.
The core solution is the MEA Cycle, which splits tasks into three independent roles:
- Manage: Can use lightweight models.
- Execute: Can use stronger models.
- Audit: Only the auditor updates task state based on independent environmental verification, serving as the sole cross-turn memory.
Significant Results: Using the same model (Qwen 3.7-Plus), simply switching to this framework increased WeaveBench pass rates from 51.8% to 80.7% and OSWorld 2.0 completion from 2.8% to 8.3% (a 3x improvement), while reducing token consumption by 24%. It supports seamless switching between GUI and CLI, with some metrics outperforming Claude Opus 4.7 running Claude Code.
More from coding & agent
- Dev Shares How Their Coding Agent Judges Them for Doom Scrolling — rudrank · 2026-08-04
- GreenPT & Open Source Tools: Compress AI Output Without Modifying Models — BaXRS1988 · 2026-08-04
- Borrowing from Browsers: New Terminal Multiplexer Introduces App Icons — evilrabbit_ · 2026-08-04
- Opinion: Why Cloud AI Agents Need Code Execution Capabilities — SnooPeripherals5313 · 2026-08-04
- Indie Dev Shares $1M MRR Tech Stack: AI Agents and Automation Tools — tibo_maker · 2026-08-04
- Open-Source AI Design Skills: Stop Cursor and Claude from Generating UI Slop — sujingshen · 2026-08-04