AIAny
Icon for item

OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents

Turns open-ended everyday requests into a managed long-horizon execution process that decomposes tasks into bounded subtasks, maintains compact execution memory under context pressure, and verifies and repairs final deliverables. Designed to run unchanged across multiple LLM backends and evaluated on AgentIF-OneDay.

Introduction

Most real-world personal and desktop requests span many heterogeneous steps, tools, and modalities, so a single uninterrupted agent loop quickly loses track of goals, state, and verification. OneDayAgent's core insight is to treat an open-ended request as a managed execution process: decompose work into bounded subtasks, keep only a compact, audited task state across rounds, and verify/repair the final deliverable before declaring success. This lets the harness tolerate different model behaviors and context limits while preserving correctness guarantees.

Key Findings
  • Managed MEA loop (Manage–Execute–Audit): each round the manager emits a bounded subtask contract, a fresh-context executor runs it, and an independent auditor inspects environment changes to update a compact, verifiable task state — so only audited facts persist across rounds.
  • Robust cross-backend generalization: the same harness runs without tuning across multiple LLM backends and execution backends, enabling consistent workflows even when models exhibit different execution styles.
  • Empirical gains on a long-horizon benchmark: when evaluated on AgentIF-OneDay across 104 tasks the harness achieves strong overall performance (reported top score 0.821 with a particular backend), demonstrating practical improvements on long, cross-environment tasks.
  • Practical tradeoff: discarding executor interaction history between rounds reduces state bloat and drift but requires reliable auditing; the harness focuses engineering effort on verification primitives and subtask contracts rather than monolithic context retention.
Who It's For and Tradeoffs

Great fit if you build or evaluate LLM-driven agents that must coordinate across tools, UIs, or terminals and need a repeatable, auditable execution pipeline. The harness is useful when model variability, context limits, or multi-step verification are primary failure modes. Look elsewhere if your tasks are short, single-step API calls or you prefer end-to-end learned controllers without explicit audited state — OneDayAgent trades end-to-end flexibility for modular correctness and auditability.

Information

  • Websitearxiv.org
  • AuthorsJingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, Ningyu Zhang
  • Published date2026/08/04

Categories

More Items

Orchestrates reasoning, external tool use, and native image generation under one unified multimodal agent policy via post-training. Introduces RAD-GRPO for agentic reinforcement fine-tuning and releases training data plus the full post-training infrastructure.

A 365-day, order-level simulation benchmark for evaluating long-term coherence of LLM agents in seller-side e-commerce. Grounded in 98,843 real product records and 26 interactive tools, it pairs prompt upstream supplier signals with delayed downstream order outcomes to stress planning, memory, and tool use over long horizons.

Frames skill generation as a sequential editing task and introduces a novel rollback reward to train an RL generator (Skill-α) that evaluates each edit by its downstream execution impact, producing skills that improve agent success rates across document-to-skill and experience-to-skill settings.