*** AgentFounder: Tongyi DeepResearch

Three days ago, Alibaba open-sourced AgentFounder, a new AI research agent that achieves state-of-the-art performance by integrating planning, tool-use, and multi-step reasoning directly into its foundational training process, rather than adding these capabilities later.

Imagine you are to predict the path of a hurricane.

Traditional AI Training: You might build a highly detailed model of the atmosphere and oceans, and then, as a separate step, try to teach it how to interpret satellite data and radar to adjust its predictions in real-time. It is good at modeling, but connecting that to real-world observations and decision-making is a secondary skill.

AgentFounder Training (Agentic CPT): It builds a hurricane prediction system that, from its earliest stages, is constantly learning to integrate diverse data sources (satellite images, buoy readings, weather balloons) and execute actions (run different simulation models, compare forecasts, identify uncertainties) to dynamically refine its predictions. It is designed from the ground up to be an active, adaptive forecaster, not just a passive data processor.

So, AgentFounder is like a dynamic, data-integrating hurricane forecaster, inherently built to interact with its environment (the research landscape) and make adaptive decisions to arrive at the best possible forecast (research insights).

A short video will be uploaded to illustrate the essence of AgentFounder.

@misc{su2025scalingagentscontinualpretraining, title={Scaling Agents via Continual Pre-training}, author={Liangcai Su and Zhen Zhang and Guangyu Li and Zhuo Chen and Chenxi Wang and Maojia Song and Xinyu Wang and Kuan Li and Jialong Wu and Xuanzhong Chen and Zile Qiao and Zhongwang Zhang and Huifeng Yin and Shihao Cai and Runnan Fang and Zhengwei Tao and Wenbiao Yin and Chenxiong Qian and Yong Jiang and Pengjun Xie and Fei Huang and Jingren Zhou}, year={2025}, eprint={2509.13310}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2509.13310}, }