# NVIDIA details session-aware routing for agent workloads in Dynamo

_Published Thursday, October 8, 2026 at 2:16 PM EDT · AI, Infrastructure · Latest · Tier 2 — Notable_

![NVIDIA details session-aware routing for agent workloads in Dynamo — Primary](https://pytorch.org/wp-content/uploads/2026/10/All-PyTorch-Blog-Social-Images-31.png)

NVIDIA says its Dynamo inference software now links an agent's repeated model requests through a shared session identifier, supporting routing and memory management across an entire task. Claude Code, Codex and OpenCode supply compatible identifiers without configuration; custom agent software can add a request header.

A native scheduling plugin tracks sessions and can pause them at tool boundaries under memory pressure. This targets repeated eviction and rebuilding of cached model context when concurrent agents exceed available memory. Session-linked traces also support simulated and live workload replay. The shared cache indexer remains experimental, and an interface for directing cache movement is proposed.

## Sources

- [PyTorch](https://pytorch.org/blog/session-aware-agentic-inference-with-nvidia-dynamo/)

---
Canonical: https://techandbusiness.org/newswire/vokjE6g0eQKl9Q2fQdPELr
Published: 2026-10-08T18:16:24.421Z
Story chronology: 2026-10-08T18:13:46.000Z
Retrieved: 2026-10-08T20:38:03.624Z
Publisher: Tech & Business (techandbusiness.org)