What changed
AMD has announced the release of PrimusClaw, an open-source managed agent runtime designed for Kubernetes clusters that host AMD Instinct™ GPUs. The tool provides a session store, an agent harness, and per-session sandboxes that run within the user's own Kubernetes environment. PrimusClaw is currently in production use at AMD, with two deployments hosting over 40,000 agent sessions for more than 70 users, accumulating over 100,000 agent-hours, more than 4 million tool calls, and over 500 billion prompt tokens.
What this means in practice
PrimusClaw allows organizations to run managed agent services alongside their AMD Instinct GPU clusters, enabling long-running tasks such as kernel tuning, inference server optimization, and extended experiments. The runtime is built to persist across upgrades and failures, with production data showing that 95.7% of runs that lost their sandbox mid-flight succeeded, and 99.5% of runs during API rolling upgrades completed successfully. The system also features an idle sandbox reclamation mechanism that only reclaims sandboxes after confirming that all associated jobs have completed. The runtime is MIT-licensed, except for the sandbox control plane, which is an Apache-2.0 fork of the Kubernetes SIG Apps agent-sandbox project. It supports standard interfaces compatible with Claude Managed Agents, A2A v1.0, and MCP, with models served through Anthropic- or OpenAI-compatible gateways. GPU-native operations are supported through ROCm™-based images, and multi-node GPU clusters can be managed via Primus-SaFE, AMD’s open-source platform.
Limitations and availability
PrimusClaw is available as an open-source repository, requiring users to operate it on their own Kubernetes clusters. Authentication and multi-node GPU placement are handled by Primus-SaFE, which is required for production installs. The runtime supports GPU optimization agents such as Hyperloom and GEAK. The supplied announcement does not specify exact hardware requirements or licensing details beyond the MIT and Apache-2.0 licenses. The tool is designed for organizations that operate their own GPU clusters and wish to run the full runtime, including the harness and session log, within their infrastructure.
