Editorial illustration for PrimusClaw Open-Source Runtime Manages AMD Instinct GPU Clusters
PrimusClaw Open-Source Runtime for AMD GPU Clusters
AMD has been running agent workloads on its own hardware since April 2026, and the numbers add up fast: more than 40,000 agent sessions, over 100,000 agent-hours, 4 million tool calls, and 500 billion prompt tokens across two internal deployments. The code behind that usage, called PrimusClaw, is now open source.
The pitch is simple. Anthropic, AWS, Google, Microsoft, and OpenAI all offer managed agent services now, running either their own harness or a customer's, plus the sandbox and session state that go with it. That works fine for a lot of use cases.
But organizations running their own AMD Instinct GPU clusters often need something different: agents that live right next to the accelerators they're tuning kernels on, optimizing inference servers for, or running multi-hour experiments against. Shipping that work out to someone else's cloud defeats the purpose.
PrimusClaw is AMD's answer to that gap, a managed agent runtime built to run inside a customer's own Kubernetes cluster rather than a vendor's. It's structured around three planes, session, harness, and sandbox, and built to keep long-running agent work alive through upgrades and failures.
PrimusClaw is an open-source managed agent runtime: a session store, an agent harness, and per-session sandboxes that you run in your own Kubernetes cluster, next to your AMD Instinct™ GPUs. In production use at AMD since April 2026, two PrimusClaw deployments have hosted more than 40,000 agent sessions for more than 70 users: over 100,000 agent-hours, more than 4 million tool calls, and more than 500 billion prompt tokens.
Why this matters
The number that stands out here isn't the 40,000 sessions or the 4 million tool calls, it's the 2,140 runs that lost their sandbox and kept going anyway. That's the actual claim AMD is making: agent infrastructure that survives rolling upgrades and node failures without losing a session's place in the work. For teams building on Instinct GPUs, this is a concrete answer to a problem most agent frameworks paper over, what happens when the thing running your long-lived agent dies mid-task.
Open-sourcing it next to the production numbers is the credible part. AMD isn't asking anyone to trust a slide deck, they're publishing a runtime that's already absorbed outages at scale.
We'd still want to see how PrimusClaw performs outside AMD's own cluster, on someone else's Kubernetes setup with different failure patterns, before calling it a standard. But for founders and researchers wiring agents into GPU infrastructure, this is a useful reference design: lease-based recovery, per-session sandboxes, and resume-from-last-turn logic, built and battle-tested, not theorized.
Related Reading
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Axiom Math raises USD 200M Series A led by Menlo to keep AI slop out of math
From Archives Hundreds of AI hacking tools left publicly exposed online
From Archives AI Reshapes San Franciscans' Dating and Emotional Habits
Common Questions Answered
What is PrimusClaw and what are its core components?
PrimusClaw is an open-source managed agent runtime designed to run in Kubernetes clusters alongside AMD Instinct GPUs. It consists of three main components: a session store, an agent harness, and per-session sandboxes that enable organizations to manage agent workloads on their own infrastructure rather than relying on third-party managed services.
How has PrimusClaw performed in AMD's internal deployments since April 2026?
AMD's two internal PrimusClaw deployments have hosted more than 40,000 agent sessions for over 70 users, accumulating over 100,000 agent-hours, 4 million tool calls, and 500 billion prompt tokens. This production usage demonstrates the runtime's capability to handle substantial agent workloads at scale.
What makes PrimusClaw's infrastructure resilience significant compared to other agent frameworks?
PrimusClaw can survive rolling upgrades and node failures without losing a session's place in the work, as evidenced by 2,140 runs that lost their sandbox but continued operating anyway. This resilience addresses a critical problem that most agent frameworks overlook: maintaining long-lived agent sessions when the underlying infrastructure fails or requires updates.
How does PrimusClaw compare to managed agent services offered by major cloud providers?
While Anthropic, AWS, Google, Microsoft, and OpenAI all offer managed agent services that run their own harness with sandboxes and session management, PrimusClaw provides an open-source alternative that allows teams to run the same infrastructure in their own Kubernetes clusters next to their AMD Instinct GPUs. This gives organizations greater control and flexibility over their agent deployment environment.
Further Reading
- PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters - ROCm Blogs
- GitHub - AMD-AGI/PrimusClaw: LLM Agent Orchestration on Kubernetes - GitHub
- Rebuilding Agentic AI from First Principles for AMD GPU - AMD Developer Resources
- Deploying OpenHands Coding Agents on AMD Instinct GPUs - AMD Developer Resources
- The Importance of Open Ecosystems for Agentic AI Deployment - AMD Blogs
