Vanishing Gradients Episode 63: Why Gemini 3 Will Change How You Build AI Agents with Ravin Kumar (Google DeepMind)
103 snips
Nov 22, 2025 Ravin Kumar, a researcher at Google DeepMind specializing in generative models and LLM products, joins to discuss the groundbreaking Gemini 3. They illustrate how models can 'self-heal' and adapt, reshaping software development. Topics include the transition from basic tool calling to advanced agent harnesses, the contrast between deterministic workflows and high-agency systems, and the importance of robust evaluation infrastructures. Ravin also shares insights on the evolution of productive features like Audio Overviews and the future of multimodal agents.
AI Snips
Chapters
Transcript
Episode notes
Models Now Do Multi-Minute Self-Repair
- Models like Gemini 3 can run multi-minute, multi-step reasoning sessions that call tools and self-correct repeatedly.
- This elevates models from short completions to self-sufficient problem solvers for complex tasks.
Three Rewrites Of Google Agent Harnesses
- Ravin rewrote internal agent harnesses three times in three years as model capabilities improved.
- He removed defensive logic when models could reliably handle tasks and rebuilt harnesses for new, more complex behaviors.
Stronger Models Support More Tools
- Better models can handle more tools in-context, reducing confusion and enabling broader multifunction agents.
- This lets a single agent handle diverse tasks instead of separate narrow bots.