PyTorch Conference NA 2026：Agentic AI

TL;DR

PyTorch Conference North America 2026 features Agentic AI and Next-Gen Intelligence across sessions on training agents, serving agents in production, agents that build PyTorch, and PyTorch in the physical world.

Agentic AI and Next-gen Intelligence at PyTorch Con NA

When you look at the schedule for PyTorch Conference North America 2026, one thing jumps out: agents are featured everywhere. They write kernels, triage CI, migrate workloads between chips, drive robots, and answer the phone. The interesting questions have shifted from whether a model can do any of these things to how we train, serve, govern, and debug systems that act on their own.

Here is a guided tour of the agentic AI and next-generation intelligence content across the two days of PyTorch Conference North America, along with why it is worth being in the room.

View the full conference schedule

Register for PyTorch Conference North America 2026

It starts on the keynote stage

Three keynotes frame the whole conversation.

Beyond Brute Force: The Era of Adaptive Intelligence

Sara Hooker, Adaption

October 20, 09:40am | Grand Ballroom

The next unlock isn’t scale, it’s architecture: systems that keep learning after deployment, closing the gap between a model’s frozen training distribution and the world it actually operates in. Sara’s talk digs into continual, gradient-free learning, which includes updating behavior without full retraining, without catastrophic forgetting, and without paying for repeated fine-tuning cycles. If you have ever shipped a model and watched it slowly drift out of relevance, you will not want to miss this.

Workload Fungibility in the Age of Agents

Bill Jia, Google Cloud

October 21, 09:15am | Grand Ballroom

The other half of the story addresses agents as developers. Alongside the deep dive on TorchTPU going open source, Jia demos long-horizon agentic workflows that migrate complex model workloads from GPUs to TPUs. The workflows keep going, hill-climbing on quantization, custom kernel generation, and sharding strategies with minimal human intervention.

Agentic AI Foundation Keynote

Mazin Gilbert, Agentic AI Foundation

October 21, 09:10 | Grand Ballroom

Mazin Gilbert is the Executive Director of the Agentic AI Foundation at the Linux Foundation and has over 25 years of experience pioneering open source platforms, authoring 100+ research papers, and holding 260+ U.S. patents. In his keynote, he will provide critical context on how agentic infrastructure is establishing a neutral governance home to support scalable, enterprise-ready open source AI deployment.

Training agents: RL becomes a requirement

The single densest cluster of agentic content is in post-training. Multi-turn, tool-using, long-horizon RL has moved from research curiosity to production requirement.

Agentic RL Training in PyTorch

Yichuan Wang, Shuhua Yu, Meta

October 20, 16:20 | 210AE

An end-to-end overview of the reinforcement learning training loop, spanning rollout infrastructure, trainer-serving communication, environment abstractions, and sandbox execution, alongside task scheduling and the trade-offs between on-policy and off-policy methods.

Open Source Reinforcement Learning with Agent Harnesses

Ben Burtenshaw, Hugging Face

October 20, 17:30 | 210AE

OpenEnv is the interoperability layer for publishing and running RL environments, co-owned by Hugging Face, Meta, Unsloth, Prime Intellect, Modal, NVIDIA, Mercor, and others. Frontier labs train models inside their own harness, while the open ecosystem vendors model, harness, and train separately. This talk is about closing that gap, and it ends with a demo you can fork on one GPU.

Train the Agent, Not Just the Model

Sergio Paniego Blanco, Hugging Face

October 21, 12:20 | LL21DEF

Explore practical SFT and GRPO techniques in agentic environments, progressing to harness-driven training where the harness manages its own inner loop directly inside the environment.

Torchtitan RL: A Unified and Extensible Training Framework for Agentic Tasks

Felipe Mello, Jiani Wang, Meta

October 21, 16:20 | LL21DEF

In Felipe and Jiani’s session, attendees learn how to use a single model definition for both training and generation with an on-policy, bitwise-reproducible mode that keeps system artifacts from getting in the way of your reward design.

Miles: Enterprise-facing Agentic RL Framework

Mao Cheng, RadixArk

October 21, 11:45 | LL21DEF

Learn how to scale post-training using unified low-precision training, stable Mixture-of-Experts (MoE) reinforcement learning, and accelerated speculative rollouts.

When Rollout and Training Disagree

Neiwen Ling (ByteDance), Tianle Zhong (University of Virginia)

October 21, 16:55 | LL21DEF

A sharp, specific lightning talk on training-inference mismatch: why small token-level numerical disagreements between the rollout engine and the training path are not benign, and how they can quietly reshape your PPO/GRPO objective.

Serving agents in production

Agentic workloads break the assumptions inference stacks were built on. Sessions may sit idle for hours before a follow-up arrives. Context is long, multi-turn, and tool-laden. CPU work such as orchestration, tool execution, and scheduling stops being a rounding error.

Making Enterprise Agentic Inference Production-Ready with PyTorch and vLLM

Joseph Groenenboom, Tyler Michael Smith, Red Hat

October 20, 15:25 | LL21ABC

Joseph shares what enterprise readiness actually entails, from build infrastructure up through tool calling and long-context multi-turn chat, with practical insights from the engineers themselves.

Sponsored: PyTorch for Agentic AI: Scaling Heterogeneous Systems from CPU to XPU

Eikan Wang, Huma Abidi, Intel

October 20, 17:30 | LL21DEF

To accommodate system-intensive and heterogeneous agentic AI workloads, Intel employs an upstream-first strategy across Intel Xeon processors, Intel GPUs, and its open software stack to deliver standard PyTorch workflows and deep performance optimizations.

Native Tiered KV Cache Offloading in vLLM

Or Ozeri, IBM

October 20, 14:50 | LL20CD

To address LLM scaling challenges with long-lived agentic sessions, vLLM introduces an upstream, dependency-free tiered KV cache offloading framework that routes transfers through CPU memory as a universal transport hub to minimize GPU overhead and ensure hardware-agnostic compatibility.

LMCache: a cluster-wide open source solution for LLM prompt caching

Kuntai Du, Tensormesh

October 20, 15:40 | LL20CD

LMCache offers a popular, open source prompt caching solution featuring extensive support across major inference engines and storage backends, alongside Kubernetes deployment guidance and underlying research insights.

Agents that build PyTorch

This is the most self-referential thread on the schedule, and one of the most practical.

Contributing to PyTorch with AI agents (Birds of a Feather)

Edward Yang, Meta

October 20, 10:35 | Community Expo

How should agents be used to contribute productively? How do you get your PR reviewed? Come, discuss and share your perspectives in person.

Fighting Agents with Agents: Bringing Claude to PyTorch CI, triage, and PR review

Driss Guessous, Meta

October 20, 12:35 | LL21DEF

Maintainers are already reviewing an ever-increasing number of agent-written PRs. This is the story of giving maintainers agent-shaped infrastructure for an agent-shaped world: issue triage, PR review skills, autorevert investigation, and the adoption curve after launch.

Shipping PyTorch and Its Ecosystem: A Modern Release Story

Andrey Talman, Meta

October 20, 11:45 | LL21DEF