AMD releases PrimusClaw open-source managed agent runtime for Instinct GPU clusters

AMD has announced the release of PrimusClaw, an open-source managed agent runtime designed to run within private Kubernetes clusters alongside AMD Instinct GPUs. The software provides a complete infrastructure for long-running AI agents, including a session store, an agent harness, and isolated per-session sandboxes. This allows organizations to host their own agentic workflows locally rather than relying on managed cloud services from providers like OpenAI or Anthropic.

The runtime is architected to be highly resilient, ensuring that agent sessions can survive hardware failures, rolling API upgrades, or lost sandbox pods. By decoupling the agent loop from specific worker processes, PrimusClaw allows sessions to be resumed from checkpoints on different replicas. The project is released under the MIT license, with the sandbox control plane being an Apache-2.0 fork of the Kubernetes SIG Apps agent-sandbox project.

Technical architecture and resilience features

• Three-Plane Architecture: The system separates operations into a session plane (Fastify/PostgreSQL for state), a harness plane (stateless 'Brain' for the agent loop), and a sandbox plane (isolated 'Hands' for code execution).
• Fault Tolerance: Agent chat turns are stored as database rows under leases rather than being pinned to active processes, enabling seamless recovery of runs during rolling upgrades or sandbox failures.
• Security and Isolation: Agent-generated code is executed in sandboxes that lack platform-wide credentials; outbound traffic is filtered, and calls must use short-lived session-bound tokens.
• GPU Integration: The runtime supports ROCm-based images and can request AMD Instinct GPUs, integrating with the Primus-SaFE cluster platform for multi-node GPU placement.
• Standardized Interfaces: It maintains compatibility with the Claude Managed Agents flow, A2A v1.0, and the Model Context Protocol (MCP) in both directions.

Why it matters

PrimusClaw enables enterprises and researchers to deploy sophisticated, long-running AI agents on their own hardware without vendor lock-in to major cloud providers. By providing a Kubernetes-native way to manage stateful agent sessions, it bridges the gap between raw GPU compute and the managed agent experiences currently dominated by closed-source APIs.