Building Tools for Agents at Weights & Biases

Watch this 42-minute keynote presentation from Fully Connected London 2025 featuring Weights & Biases Co-founder Lukas Biewald and CoreWeave VP of Engineering Camille Fournier as they explore the challenges of AI roadmaps and demonstrate cutting-edge tools for building AI agents. Discover why creating roadmaps in AI development is particularly difficult and gain insights into the latest developments from both Weights & Biases and CoreWeave over the past six months. Learn about W&B Models and W&B Weave, two key platforms that have been developed to support AI workflows, and explore the collaboration between Weights & Biases and CoreWeave in advancing AI infrastructure. Examine CoreWeave's AI object storage solutions and understand how W&B Inference integrates with Marimo for enhanced AI development workflows. Follow a detailed walkthrough of building a Software Engineering (SWE) Agent at Weights & Biases, including comprehensive agent architecture and design principles. Explore the experimentation phase of agent building and investigate whether AI can automatically improve experimental loops for more efficient development cycles.

Syllabus

00:00 - Introduction and why roadmaps are hard
00:57 - W&B Models: What we’ve been building over the last three months
07:16 - W&B Weave: What we’ve been building over the last three months
11:50 - Introducing Camille and what Weights & Biases and how CoreWeave are collaborating
18:22 - Walking through CoreWeave’s AI object storage
21:20 - W&B Inference and Marimo
25:40 - How we build a SWE Agent at Weights & Biases and agent overview
30:45 - Experimentation phase of agent building
35:15 - Can AI automatically improve the experimental loop?