CognitiveKernel-Pro is a state-of-the-art open-source deep research agent that uses free tools (with optional Google Search API). The project provides fully reproducible supervised fine-tuning recipes for training agent foundation models without requiring reinforcement learning, along with RL training recipes (subgoal-GRPO) for further improvements. It publishes datasets on Hugging Face for both query data and SFT training, as well as trained model weights (Qwen3-8B-CK-Pro). The associated research has been accepted at ACL 2026.