PyTorch-Native Stack for Agents - Allen Wang & Davide Testuggine, Meta

Description

PyTorch-Native Stack for Agents - Allen Wang & Davide Testuggine, Meta Agents are the next frontier of AI development — systems that reason, act, and learn through interaction. In this talk, we’ll explore the challenges of building for agents and the new infrastructure they demand, starting with a practical introduction to reinforcement learning. We’ll then unveil the PyTorch-native stack we’ve built to empower this new era of development. TorchForge — a PyTorch-native library for scalable RL post-training and agentic development, designed to let researchers focus on algorithms rather than infrastructure. TorchStore — a high-performance tensor key-value store optimized for RDMA-enabled environments while maintaining the simplicity and usability of PyTorch. OpenEnvironments — an open initiative to build and share reinforcement learning environments collaboratively with the community. Together, these projects form the foundation for the next generation of agentic AI development — from research to production.