Hugging Face researchers have introduced OneDayAgent, a new harness designed to help LLM agents handle open-ended, long-horizon tasks that span work, study, and daily life. These tasks are inherently complex, requiring agents to maintain goals and constraints across many steps while navigating heterogeneous tools and multimodal attachments.
Previous research has tackled individual failure modes such as goal drift, state loss, and context overflow, but a unified solution that addresses them jointly and works across different backends has been less explored. OneDayAgent aims to fill this gap by turning an open-ended request into a managed execution process.
The harness decomposes tasks into bounded subtasks, maintains execution memory under context pressure, and verifies and repairs the final deliverable. This structured approach ensures that agents can handle long-horizon tasks without losing track of objectives or running into context limitations.
In evaluations on the AgentIF-OneDay benchmark, which includes 104 tasks, OneDayAgent achieved a state-of-the-art overall score of 0.821 when paired with the GLM-5.2 backend. Notably, the same harness ran successfully across five backend LLMs from three different model families, demonstrating its ability to generalize without tuning, even as different models exhibit distinct execution styles under the same workflow.
The researchers emphasize the potential of OneDayAgent to enable more robust and versatile autonomous agents, capable of handling complex, real-world requests that require long-term planning and execution.