PyTorchCon 2026 Keynote Maps Five Layers of the Agentic Stack
The bidirectional design idea turns agent evaluation data into training rewards, a concrete recipe for closing the loop between deployment and model improvement. Researchers and framework builders can experiment with reward signals derived from real agent traces rather than static benchmarks. The framing also clarifies where projects like PyTorch and the Agentic AI Foundation sit relative to each other in the stack.
At PyTorchCon 2026, Mazin Gilbert, Executive Director of the Agentic AI Foundation, presented a framework for how LLMs can learn from agent behaviour. He mapped the agentic stack into five layers: hardware infrastructure, software infrastructure, intelligence, the agentic platform, and user applications, placing PyTorch at the intelligence layer and the Agentic AI Foundation at the platform layer. Gilbert proposed designing the stack bidirectionally, so agent traces, evaluations, environments, and rewards feed back into model training. As an open research direction, he described converting agent evaluations into training rewards in PyTorch, covering safety, task completion, precision, efficiency, and energy use.
How can LLMs learn from what agents do? At #PyTorchCon 2026, Mazin Gilbert, Executive Director of @AgenticAIFdn, examines the relationship between models and agents.
Mazin maps five layers of the agentic stack: hardware infrastructure, software infrastructure, intelligence, the agentic platform, and user applications. He places PyTorch at the intelligence layer and Agentic AI Foundation at the agentic platform layer.
He proposes designing the stack in both directions: models support agents, while agent traces, evaluations, environments, and rewards feed back into how models are built.
As an open research direction, Mazin describes turning agent evaluations into rewards for training in PyTorch, accounting for safety, task completion, precision, efficiency, and energy use.
Watch the keynote livestream: https://t.co/6TtaEkvsHz
