TL;DR: Three failure modes of long-horizon agents — compounding errors, context rot, and task-state loss — solved structurally through a Manage-Execute-Audit (MEA) loop. WeaveBench PassRate: 51.8% → 80.7%. LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks Key points: 🔧 Manager: Maintains explicit task state outside the execution trajectory. Constructs subtask contracts specifying goals, acceptance criteria, constraints, and evidence. ⚡ Executor: Runs each subtask in a fresh, budget-bounded context — no prior trajectory history passed in. 🔍 Auditor: Post-execution read-only inspection, independent of the executor. Reports completion, integrity, and state updates; serves as persistent cross-round memory. 🖥️ GUI/CLI hybrid: Manager routes tasks to the right interface; AgentAdapter swaps in Claude Code, Codex CLI, Hermes Agent without modifying native loops. 📊 WeaveBench: PassRate 51.8% → 80.7%; Design +60pp, Spatial/3D +50pp. 🤖 OSWorld 2.0: Qwen 3.7-Plus 2.8% → 8.3% (3×); Claude Opus 4.7 20.6% → 35.3%. 💻 Terminal-Bench: 69.7% → 77.2%, with 24% fewer tokens consumed. 💡 Manager overhead: only 2–8% of total tokens. Auditor carries 19–38% — the primary cost of reliability. "Agent capability is a property of the complete model–harness system, not the model alone" — that framing changes how you should think about long-horizon agent design. #AIAgents# #LLM#