[research] Fresh-context execution loops cut agent failures by ~36 points on long tasks

[github-actions[bot]]

🔬 The Finding

Researchers introduced LongHorizon-Harness (arXiv, Aug 4 2026), a harness that separates task-state management from task execution. Instead of letting a growing context track progress, a dedicated manager maintains verified state, a fresh-context executor handles each subtask, and a read-only auditor checks results before the next step. This Manage-Execute-Audit (MEA) loop lifted Qwen 3.7-Plus from 51.8% → 80.7% on WeaveBench and boosted Claude Opus 4.7 from 20.0% → 34.3% on OSWorld 2.0.

⚙️ What It Means for Agentic Workflows

• Don't let agents track their own progress in-context. Externalizing task state and verifying it independently before each subtask eliminates compounding self-assessment errors — a key failure mode in multi-step automation pipelines.
• Fresh-context executors are a drop-in pattern. The lightweight AgentAdapter works with existing model and harness backends unchanged, so you can retrofit this pattern into current workflows without rewriting agent logic.

🔗 Source

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks — August 4, 2026

Generated by Daily Agentic AI Research Digest · 55.4 AIC · ⌖ 11.5 AIC · ⊞ 24.2K · ◷

• expires on Aug 12, 2026, 10:29 AM UTC