Long Horizon Harness

LongHorizon-Harness (LH harness) is a runtime for long-running agent tasks built around a Manage-Execute-Audit (MEA) loop. Instead of one growing session that judges its own work, it separates three roles: a manager that keeps task state and plans the next subtask, an executor that works a single subtask in a fresh context, and a read-only auditor that inspects the real environment and writes the audit report.

Task state is written only from audit evidence, so an item counts as completed only when a report supports it. The executor's raw trajectory is discarded after each round and only compact audit reports carry over. The aim is to reduce compounding errors, context rot and task-state loss. If the auditor modifies protected state, its report is flagged as an integrity violation.

A lightweight AgentAdapter lets Claude Code, Codex CLI, OpenClaw and Hermes Agent act as interchangeable backends for any role, with models such as Claude Opus, GPT and Qwen. Budgets are 1800s per round for the executor and 300s for the manager and auditor, with a cap of 25 MEA rounds per task. The site reports 80.7% PassRate on WeaveBench, 35.2% partial score on OSWorld 2.0 and 77.2% success on Terminal-Bench 2.1. A YouTube reviewer notes the gains come at some extra token cost. The vendor, license and source repository were not identified in the provided material.