Agentic AI and Next-Gen Intelligence at PyTorch Conference 2026
Key sessions focusing on training and deploying AI agents in production.
1Conference Main Theme
PyTorch Conference North America 2026 focuses on Agentic AI and next-gen intelligence systems. Topics include agents writing core software, managing CI processes, and optimizing workloads across chips. The question is no longer whether models can do these tasks, but how to train, serve, and govern them.
2Keynote Speeches
Three major keynotes set the conference's framework. Sara Hooker discusses systems enabling continuous post-deployment learning in 'Beyond Brute Force: The Era of Adaptive Intelligence.' Bill Jia from Google Cloud covers TorchTPU going open source and agent workload migration from GPUs to TPUs.
3Agentic AI Foundation
Mazin Gilbert, Executive Director of the Agentic AI Foundation, explains how neutral governance structures support open-source AI deployment. With 25 years of experience and 260+ patents, he sheds light on the future of agent infrastructure.
4RL Training
Agent training has evolved from research curiosity to production necessity. Yichuan Wang and Shuhua Yu from Meta detail the RL training loop, while Ben Burtenshaw from Hugging Face demonstrates open-source RL environments in action.
5Open Source RL
OpenEnv, developed by Hugging Face, Meta, and others, serves as an RL environment layer. While frontier labs train models in proprietary environments, the open ecosystem separates models, environments, and training. This session shows how to bridge that gap.
6Practical SFT and GRPO
Sergio Paniego Blanco from Hugging Face explains practical SFT and GRPO techniques in agent environments. He demonstrates how to build systems where the environment manages its own inner loop.
7Torchtitan RL Framework
Felipe Mello and Jiani Wang from Meta show how to use a single model definition for both training and generation. Their bitwise-reproducible mode prevents system artifacts from interfering with reward design.
8Miles Framework
Mao Cheng from RadixArk demonstrates scaling training using unified low-precision training, stable MoE RL, and accelerated speculative rollouts.
9Training-Inference Mismatch
Neiwen Ling and Tianle Zhong explain why minor numerical discrepancies between training and inference aren't benign and how they can silently reshape PPO/GRPO objectives.
10Agent Serving in Production
Agent workloads break traditional inference stack assumptions. Sessions may idle for hours, contexts grow long and multi-turn, and CPU work like orchestration becomes significant.
11Enterprise Agent Inference
Joseph Groenenboom and Tyler Michael Smith from Red Hat detail what enterprise readiness entails, sharing practical insights on tool calling and long-context multi-turn chat.
12Intel's Approach
Intel adopts an upstream-first strategy for heterogeneous agent AI workloads, leveraging Xeon processors, Intel GPUs, and open software stacks to deliver standard PyTorch workflows with deep optimizations.
Original article: PyTorch Blog — PyTorch Konferansı 2026'da Agentic AI ve Yeni Nesil Zeka