Reproduce “OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, Ningyu Zhang
OneDayAgent is an autonomous agent harness that structures open-ended, long-horizon tasks, addressing challenges like goal drift and context overflow through integrated task decomposition, execution memory, and global verification. It achieved an overall score of 0.821 on the AgentIF-OneDay benchmark and demonstrated generalizability across diverse large language model backends.
Key insights
- Explicitly structuring long-horizon tasks through decomposition into subtasks and incorporating artifact-level global verification significantly enhances agent reliability and deliverable compliance.
- Effective execution memory, which includes summarized truncation and automatic context compression, allows agents to manage high context pressure without degrading task quality.
- The harness design is generalizable, allowing it to operate stably and effectively across various LLM backends from different model families without requiring backend-specific tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.