Cracking open an open model | The Agent Factory Podcast

How do you actually build and train an open source AI model from scratch? This episode of The Agent Factory cracks open the entire "factory line" for creating new agentic models. Forget just using APIs; we're talking about the data, the supervised fine tuning, the reinforcement learning, and the evals—the whole process. ⚙️ Join hosts Ivan Nardini and Amit Maraj, along with special guest Ravin Kumar from the DeepMind team, as they unpack the complete development lifecycle for building your own open source agents. In this episode you'll learn: 1️⃣ Agent Industry Pulse: We cover a packed couple of weeks in AI news. We explore the new Gemini 2.5 Computer Use model that gives AI a "virtual mouse and keyboard," the new "vibe coding" and annotation mode features in AI Studio, DeepSeek-OCR's new approach to context compression, and the massive update to Veo 3.1 that allows generative video editing. 2️⃣ The Factory Floor: The heart of this episode is the deep dive with Ravin Kumar. We get his hands on perspective on the end to end process for training an open model with agentic capabilities, including: ➖ The data: What does "agent training data" (trajectories, thoughts, and actions) actually look like? ➖ The training: A breakdown of the two phase punch: Supervised Fine-Tuning (SFT) to learn "textbook" examples and Reinforcement Learning (RL) to "learn by doing." ➖ The evaluation: How do you build the "final exam" for an agent to test for both capability and safety? Developer Q&A: We answer this top community question: ➖ "What are your recommended strategies for evaluating the performance and reliability of AI agents before deploying them to production, especially when using models like Gemini?" Whether you're a seasoned developer or just starting your journey with AI agents, this episode is packed with inspiration and practical demos that will make you want to start building immediately. Leave your questions and feedback in the YouTube comments below! 🔗 Resources