Alibaba's AMAP-ML released LongHorizon-Harness for long computer-use agent tasks

Alibaba's AMAP machine learning group has released LongHorizon-Harness, an MIT-licensed system that wraps existing coding-agent CLIs in a three-role loop designed for long-running desktop and terminal work. The repository was published this month and has gained around 1,400 stars on GitHub.

The harness runs with a CLI the team already uses: OpenAI's codex, Anthropic's claude, opencode, or DeepSeek Harness. It adds a coordination layer around whichever one is installed.

The three roles

Every task cycles through a Manager that plans, an Executor that acts on the plan, and an Auditor that verifies the result before the loop continues. If the Auditor rejects the work, the Manager replans from the last verified state. AMAP-ML describes the goal as letting agents "operate the whole computer, preserve verified progress, keep working until the task is done."

The recovery point is the interesting part. A long computer-use task that fails in step forty keeps the first thirty-nine, and the team's benchmark numbers depend on that.

The benchmark numbers

AMAP-ML reports its harness raising the same underlying agent from around 50 to around 80 on WeaveBench, giving roughly a three times improvement on OSWorld 2.0, and moving Terminal-Bench 2.1 from 69.7 to 77.2. The paper is on arXiv as 2608.01964.

Every one of these numbers is measured by the authors on their own harness. The comparison base is the same CLI without the loop, so the numbers capture what the loop itself contributes. An independent run on a different model remains the check that matters, and the MIT licence and open repository make that a one-command test.

Requirements

To try it a team needs Python 3.10, one of the four supported agent CLIs installed and configured, and Node.js 20 for the computer-use plugins that control a real desktop. A React and FastAPI dashboard is bundled for browser-based task management, and recent updates added conversation continuation after a run and phase-one integration with DeepSeek Harness.

The gap between a coding agent that can write a function and a coding agent that can finish a multi-hour desktop task is mostly loss of state. Runs stop, tool calls fail, and the agent restarts and meets the same problem it met an hour earlier. A wrapper that keeps a verified checkpoint after every step and replans from there is a plausible fix for that class of failure, and it works with the same CLI a team already trusts.

The drawback is that a three-role loop costs more tokens than a single-agent run, so the harness is worth using only when the task really is long. On short prompts a single agent will be faster and just as accurate. For a build-and-deploy chore that keeps timing out, or a research task that needs the same twenty pages read carefully, this is the approach worth trying.

Source

• AMAP-ML/LongHorizon-Harness on GitHub