LongHorizon-Harness: only audited facts cross the round boundary
Aug 3 Ma et al.: Manage–Execute–Audit keeps task state outside the growing session; fresh-context executors, read-only auditors, and Claude Code/Codex backends — WeaveBench 51.8%→80.7% with the same model.
Source: LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
Ma, Huang, Zou, Wang and coauthors (3 Aug 2026; code) locate long-horizon failure in the harness, not only the model: existing coding harnesses keep execution, task state, and completion judgment inside one growing context, so wrong self-assessments become premises for later rounds. Their fix is LongHorizon-Harness — reformulate the job as task-state management and advance state only with facts independently verified from the environment.
Each round is Manage–Execute–Audit (MEA). A manager (no environment tools) reads audited state and emits one bounded subtask contract — goal, acceptance criteria, constraints, relevant prior evidence — or done / blocked / ask. A fresh-context executor (Claude Code, Codex CLI, OpenClaw, … via AgentAdapter) performs only that contract; its raw trajectory is discarded. A read-only auditor inspects the resulting environment, returns complete/incomplete/blocked plus integrity (clean/suspect/violation), and may propose state updates the manager alone applies. Executor claims never write the persistent state.
Matched receipts with Qwen 3.7-Plus and Claude Code as backend: WeaveBench PassRate 51.8% → 80.7%; Terminal-Bench 2.1 69.7% → 77.2%; OSWorld 2.0 binary completion 2.8% → 8.3%. Claude Opus 4.7 on an OSWorld subset: 20.6% → 35.3% binary. Case studies show the mechanism: stalled GUI retries become pending requirements for later executors; visually plausible “done” fails when the auditor checks underlying styles; pre-repair evidence stays pending until captured. Token cost is not a fixed tax — auditor spend is the main adder; on Terminal-Bench LH-Harness used 24% fewer tokens at higher success.
Why it matters for code review
Always-on momentum compounds only if what crossed the night was audited environment state, not a chat transcript that graded itself. MEA is the research shape of the independent auditor pattern: separate the writer from the checker, keep only verified facts between sessions, and treat “done” as a manager decision over audit reports — the review unit for continuous factories under volume.