AMD has open-sourced PrimusClaw, a managed agent runtime for AMD Instinct GPU clusters. The announcement was published on November 6, 2026 on the official ROCm blog, and the code is available in the AMD-AGI/PrimusClaw repository on GitHub. The idea is to give organizations that run their own GPU clusters the same managed experience Anthropic, AWS, Google, Microsoft and OpenAI offer as a service — but inside the customer's own infrastructure.
What PrimusClaw is
PrimusClaw splits an agent run into three parts: a durable session, a stateless harness that runs the agent loop (the “Brain”), and a per-session sandbox that executes the agent's actions (the “Hands”). Everything runs in the customer's own Kubernetes cluster, next to the Instinct GPUs. Each run becomes a row in a PostgreSQL database, and harness replicas can be upgraded or lost mid-run without abandoning the work: another replica picks it up from a checkpoint.
Already proven in production at AMD
According to AMD, the system has been in production at the company since April 2026. Two deployments have hosted more than 40,000 agent sessions for more than 70 users, totalling over 100,000 agent-hours, more than 4 million tool calls and over 500 billion prompt tokens. On the main deployment, the median run lasted 3.4 hours.
Resilience was measured in numbers: of 2,140 runs that lost their sandbox mid-flight, 2,047 (95.7%) still succeeded. Across three rolling upgrades of the API, with 583 runs in flight, 580 (99.5%) ran to completion and 544 (93.3%) succeeded — the remaining failures were agent or sandbox errors, not a lost worker.
Open standards and a GPU focus
PrimusClaw speaks the protocols the agent ecosystem already uses: A2A 1.0, MCP in both directions — it connects to MCP servers and also exposes its own operations as MCP tools — and a gateway compatible with the Anthropic and OpenAI APIs. Authentication and multi-node GPU placement are delegated to Primus-SaFE, AMD's open platform for Instinct clusters. Sandboxes can request Instinct GPUs and start from ROCm-based images, and the runtime already hosts GPU optimization agents such as Hyperloom and GEAK.
Security: credentials stay out of the sandbox
The design assumes that model-generated code may be hostile. Platform credentials therefore stay with the harness and are not passed to the sandbox — the only credential inside a sandbox is the calling user's own key. With the hardening settings from the security guide, the sandbox also filters outbound traffic and requires a short-lived token bound to the session on every call.
The repository is MIT-licensed, with one exception: the sandbox control plane is an Apache-2.0 fork of the Kubernetes SIG Apps agent-sandbox project. The release reinforces AMD's strategy of competing with Nvidia not only on hardware, but on open software for AI data centers.
Sources: ROCm Blogs — PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters; GitHub — AMD-AGI/PrimusClaw.
Image credit: AMD / ROCm Blogs.
Transparency: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources from the internet. We recommend consulting the original sources for more details.
Muse
