AMD Open-Sources PrimusClaw: A Self-Hosted Managed Agent Runtime for Instinct GPUs
AMD has open-sourced PrimusClaw, the runtime environment that has powered its internal agents since April on its Instinct GPU clusters: durable sessions, a stateless agent loop, per-session sandboxes, all running on Kubernetes. Over 40,000 sessions and 100,000 agent hours in production, according to AMD.
Running an agent for hours no longer happens inside a terminal: Anthropic, AWS, Google, Microsoft, and OpenAI now rent environments that host the agent loop, its sandbox, and its session memory. AMD is taking the opposite approach. On November 6, the company introduced PrimusClaw on its ROCm blog and released it as open source: the same kind of environment, but installed on your own Kubernetes cluster, right next to your Instinct GPUs.
The Brain, the Hands, and the Log
PrimusClaw divides an agent run into three components. A durable session, stored in PostgreSQL; a stateless agent loop, dubbed Brain, which calls models and chooses tools; and a sandbox per session, the Hands, where commands are executed. Because each run is a leased database row, any replica can pick it up: during an update, old replicas stop taking new work, write a checkpoint, and hand off execution.
On the interface front, the tool speaks current standards: an API compatible with the Claude Managed Agents onboarding path, A2A v1.0, and bidirectional MCP, with models called via an Anthropic- or OpenAI-compatible gateway. It can therefore connect to commercial models or locally hosted open models. Sandboxes can request Instinct GPUs—its original use case being agents that spend hours optimizing compute kernels or inference servers.
Six Months in Production at AMD
According to AMD, two internal deployments handled more than 40,000 sessions, over 100,000 agent hours, and more than 4 million tool calls between April and October. Runs are long: a median of 3.4 hours, reaching up to 50.1 hours. Out of 2,140 runs that lost their sandbox along the way, 95.7% still completed successfully; during three API rolling updates, 99.5% of in-progress runs finished.
What You Have to Enable Yourself
AMD is explicit on one point: maximum security does not come out of the box. Sandboxes run with runc, which shares the host machine’s kernel. Virtual machine isolation (Kata Containers), egress traffic filtering, and session-bound tokens must be enabled before exposing a deployment, and production measurements were conducted without Kata or an egress proxy. A production setup also requires Primus-SaFE, AMD’s cluster platform.
Our take
PrimusClaw primarily serves AMD’s strategy: making its GPUs more attractive by wrapping them in the software that major cloud providers sell as a managed service. For a team that already operates its own clusters and refuses to hand over prompts, tool outputs, and code to a third-party service, this is a serious, well-documented option backed by production metrics rarely disclosed in this space. The tradeoff is openly acknowledged by AMD: self-hosting means shouldering compliance, availability, and hardening entirely on your own—the very things customers usually pay a vendor to handle. As for the production metrics, they come exclusively from AMD.
Sources
- AMD ROCm Blogs — PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters (November 6, 2026) Primary source
- GitHub — AMD-AGI/PrimusClaw Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
