AMD open-sources PrimusClaw, a managed agent runtime for Instinct GPU clusters

AMD announced on its ROCm blog on 6 November that it has open-sourced PrimusClaw, the managed agent runtime it uses internally, with the code now on GitHub (AMD-AGI/PrimusClaw). Built for AMD Instinct GPU clusters, it lets enterprises run AI agents inside their own Kubernetes clusters, next to the GPUs, instead of handing agents to a third-party cloud host.

Already proven in production

According to AMD, PrimusClaw is not an experiment: it has run in AMD production since April 2026. Two deployments have served more than 40,000 agent sessions for more than 70 users, totalling over 100,000 hours of agent runtime, more than 4 million tool calls and more than 500 billion prompt tokens.

The platform isolates each agent session in its own sandbox and centrally manages session storage and the agent harness. The design lets many agents run long, autonomous tasks while staying isolated from one another and easy to trace and debug.

A2A and two-way MCP support

Interoperability is another focus. PrimusClaw supports A2A v1.0, the agent-to-agent protocol, so incoming agent tasks become real sessions on the platform, and MCP in both directions: it can connect to external MCP servers for tools and data, and can also expose its own operations as MCP tools.

A gateway compatible with Anthropic- and OpenAI-style APIs lets teams use their own hosted models. AMD says open-sourcing PrimusClaw aims to let developers building agent systems on Instinct GPUs reuse the runtime infrastructure AMD has already validated internally, rather than building it from scratch.

Sources: AMD ROCm blog (6 November 2026); DEV Community.