Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents
Yueqi Song, Ketan Ramaneti, Zaid Sheikh, Ziru Chen, Boyu Gou, Tianbao Xie, Yiheng Xu, Danyang Zhang, Apurva Gandhi, Fan Yang, Joseph Liu, Tianyue Ou, Zhihao Yuan, Frank Xu, Shuyan Zhou, Xingyao Wang, Xiang Yue, Tao Yu, Huan Sun, Yu Su, Graham Neubig
2025-10-29
Summary
This paper addresses the difficulty of training AI agents because of how scattered and differently formatted the data used for training them is. They introduce a new standard way to organize this data to make it easier to use for training.
What's the problem?
Training AI agents requires a lot of data showing them how to perform tasks. However, this data isn't all in one place; it's spread across many different sources, each with its own unique format and way of being accessed. This makes it really hard to combine all this useful information and actually train an agent effectively, even though plenty of data exists. It's like trying to build with LEGOs when all the pieces come in different sized boxes and some require special tools to open.
What's the solution?
The researchers created something called the Agent Data Protocol, or ADP. Think of ADP as a universal translator for agent training data. It’s a simple, standardized format that can represent all sorts of tasks an agent might do, like using tools, browsing the internet, writing code, or just generally following instructions. They took 13 existing datasets and converted them all into ADP format, then showed that they could use this standardized data to train agents that performed significantly better – around 20% improvement – than agents trained without it. They also made all their code and data publicly available.
Why it matters?
This work is important because it lowers the barrier to creating better AI agents. By providing a common data format, it makes it easier for researchers and developers to share data, build training pipelines, and ultimately create more capable and reliable AI systems. It promotes collaboration and reproducibility in the field, meaning others can build upon their work and verify their results.
Abstract
Public research results on large-scale supervised finetuning of AI agents remain relatively rare, since the collection of agent training data presents unique challenges. In this work, we argue that the bottleneck is not a lack of underlying data sources, but that a large variety of data is fragmented across heterogeneous formats, tools, and interfaces. To this end, we introduce the agent data protocol (ADP), a light-weight representation language that serves as an "interlingua" between agent datasets in diverse formats and unified agent training pipelines downstream. The design of ADP is expressive enough to capture a large variety of tasks, including API/tool use, browsing, coding, software engineering, and general agentic workflows, while remaining simple to parse and train on without engineering at a per-dataset level. In experiments, we unified a broad collection of 13 existing agent training datasets into ADP format, and converted the standardized ADP data into training-ready formats for multiple agent frameworks. We performed SFT on these data, and demonstrated an average performance gain of ~20% over corresponding base models, and delivers state-of-the-art or near-SOTA performance on standard coding, browsing, tool use, and research benchmarks, without domain-specific tuning. All code and data are released publicly, in the hope that ADP could help lower the barrier to standardized, scalable, and reproducible agent training.