Overcoming State Loss in Long-Horizon Agents: New Framework Boosts Accuracy
Ziyu Ma · hf · 2026-08-04
Existing LLM agents often struggle with state tracking and error propagation due to ever-growing contexts during long-horizon tasks. Researchers introduce LongHorizon-Harness, reformulating execution as a task-state management problem.
At its core is the Manage-Execute-Audit (MEA) loop:
- Manager: Explicitly maintains task state and determines the next subtask.
- Executor: Performs subtasks using a fresh context.
- Auditor: Verifies the environment state in read-only mode before proceeding.
Experiments show significant accuracy improvements across benchmarks like WeaveBench, Terminal-Bench, and OSWorld. For instance, it raises Qwen3.7-Plus from 51.8% to 80.7% on WeaveBench and lifts Claude Opus 4.7 from 20.0% to 34.3% on an OSWorld subset.
More from coding & agent
- Dev Builds Free 'Radar' Scanner to Catch Security Gaps in AI-Generated Websites — indishere · 2026-08-04
- Include Standard Libraries in RL Environments to Stop AI Agents Reinventing the Wheel — cephaloform · 2026-08-04
- From GANs to Agent Swarms: An AI Vet Reencounters 'Adversarial Collapse' 12 Years Later — bingxu_ · 2026-08-04
- Building Social Media Agents: API Approvals Are the Real Bottleneck — arslan2012 · 2026-08-04
- Qwen 3.8 Coding Test: Nearly Matches K3 at Half the Price — bindureddy · 2026-08-04
- Skip Docker: db-here Enables Zero-Risk Database Isolation for AI Agents — andersonbcdefg · 2026-08-04