AMD opens the managed agent runtime behind its own Instinct GPU clusters

AMD opens the managed agent runtime behind its own Instinct GPU clusters

📅 November 7, 2026 • AMD ROCm blog

AMD opened PrimusClaw
AMD published a post in the ROCm blog on November 6, 2026 titled PrimusClaw: An Open-Source Managed Agent Runtime for AMD Instinct GPU Clusters, with the code in the github.com/AMD-AGI/PrimusClaw repository. PrimusClaw is a managed runtime for agents. It keeps session storage, an agent harness and separate sandboxes per session, and all of it runs in a Kubernetes cluster owned by the customer, next to AMD Instinct GPUs.

Production numbers from AMD
Per AMD, the system has run in AMD production since April 2026. Two deployments served more than 40,000 agent sessions for more than 70 users, over 100,000 hours of agent runtime, more than 4 million tool calls and more than 500 billion prompt tokens. On the main deployment, the median run duration was 3.4 hours.

Architecture and licensing
The repository is MIT licensed, with one exception: the sandbox control plane is a fork of the Apache-2.0 agent-sandbox project from Kubernetes SIG Apps. The design has three parts. A long-lived session holds the state, a stateless harness called Brain runs the agent loop and routes tools, and Hands is an MCP tool server that runs inside the sandbox of each session.

Protocols and cluster integration
PrimusClaw supports A2A v1.0, so incoming agent tasks become real sessions, and MCP in both directions, meaning the platform connects to MCP servers and also exposes its own operations as MCP tools. A gateway to an Anthropic- or OpenAI-compatible API lets teams use their own hosted models. Authentication and multi-node GPU placement are delegated to Primus-SaFE, the open AMD platform for Instinct clusters.

State, isolation and what it means for engineers
Sessions, events, memory, skills, plugins and task graphs are stored in PostgreSQL, the event stream runs over SSE with resumable cursors, and state survives a harness replica restart. Sandboxes are isolated with runc by default, Kata Containers are available as an option with a virtual machine, and a sandbox can request AMD Instinct GPUs and start from ROCm based images. For engineers, the agent runtime stops being a proprietary platform and becomes a component you deploy in your own cluster, together with control over data and models.

IMPACT A managed agent runtime that ships as source code and runs inside a cluster owned by the customer changes the procurement conversation, because the harness, the sandbox layer and the model gateway can sit next to the GPUs with the data and the models under your own control.

Sources: rocm.blogs.amd.com - github.com - aidailypost.com