DeepSeek Harness Open‑Source and Alibaba’s LongHorizon‑Harness: MEA Loop Boosts Long‑Horizon AI Agents

Overview

The DeepSeek Harness project was open‑sourced and quickly attracted 102k stars. Alibaba released a companion project, LongHorizon‑Harness, which can run continuously for dozens of hours on desktop or terminal tasks by giving a single goal to Claude Code, Codex, or DeepSeek Harness.

Manage‑Execute‑Audit (MEA) Loop

LongHorizon‑Harness replaces a single ever‑growing context with an explicit three‑stage cycle: Manage → Execute → Audit. The loop treats long‑horizon execution as a task‑state‑management problem rather than a continuously expanding conversation.

Manager : Reads the current task state S_i and previous audit report V_{i‑1}, constructs a bounded sub‑task contract c_i containing goal, acceptance criteria, constraints, and historical evidence.

Executor : Runs the contract in a brand‑new context, transforming the environment from e_{i‑1} to e_i and returning an execution report o_i. Each round is an isolated, budgeted episode; all intermediate reasoning is discarded after the report.

Auditor : With read‑only permissions, independently inspects e_i and produces an audit report v_i. It can only rely on evidence directly obtained from the environment; it cannot modify protected artifacts.

The Manager updates the persistent task state based on v_i and decides the next action: continue, finish, block, or ask the user.

Experiments

Benchmarks were run with the same underlying model but different harnesses, showing dramatic gains:

WeaveBench: success rate rose from 51.8% to 80.7% (near‑doubling the previous best record).

OSWorld 2.0: completion rate increased from 2.8% to 8.3% (≈3×).

Terminal‑Bench 2.1: success grew from 69.7% to 77.2%; switching the backend to Codex pushed it to 83.1%.

Across eight benchmark domains, improvements were consistent, indicating that the framework augments model capability rather than merely compensating for weak models.

Analysis

Token cost analysis shows that the Manager consumes only 2.0–2.8% of tokens, while the Auditor accounts for 19–38%, the main added overhead. The average output tokens per task grew from 28.9 K to 104 K when using LongHorizon‑Harness with Qwen 3.7‑Plus, pushing the model to a stronger performance frontier.

Tasks that benefit most are those requiring preservation, inspection, and revision of inter‑dependent environment states (e.g., Design, Spatial/3D, interactive tutorials). Tasks limited by single‑step abilities (visual perception, math reasoning) see smaller gains.

Agent capability emerges as a product of “model × harness”: the harness determines how reliably per‑round results are retained and reused.

Case Studies

Case 1 – Recovering from a stuck WebRTC interaction (WEB_task_16) : The baseline retried 400+ steps after a non‑responsive “Decode As” dialog, achieving a score of 0.59. LongHorizon‑Harness recorded the failure and missing evidence in the task state; the next Executor directly targeted the missing chart and packet evidence, raising the score to 0.92.

Case 2 – Detecting a false‑positive completion in document styling (DOC_task_2) : The baseline edited the XML directly, seeing a correct title visually and stopping with a score of 0.00 because the required LibreOffice workflow was bypassed. LongHorizon‑Harness followed the GUI workflow; the Auditor parsed the XML of all 15 titles and verified the underlying style, achieving a score of 0.89. The Executor’s self‑declaration was blocked by the independent audit.

References

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
https://arxiv.org/abs/2608.01964
Github: https://github.com/AMAP-ML/LongHorizon-Harness
Website: https://lh-harness.pages.dev