AMD has open-sourced PrimusClaw, the managed agent runtime it runs on its own Instinct GPU clusters. It is a session store, an agent harness, and per-session sandboxes that run inside your own Kubernetes cluster, next to the GPUs, rather than in a vendor’s managed service.
PrimusClaw has been in production at AMD since April 2026. Across two internal deployments it has hosted more than 40,000 agent sessions for over 70 users, totalling more than 100,000 agent-hours, four million tool calls, and over 500 billion prompt tokens. The median run lasted 3.4 hours, and the longest ran for just over 50 hours.
Runs are built to survive failure: 2,140 runs lost their sandbox mid-flight, and 95.7% of them succeeded after rebuilding. Rolling upgrades are designed the same way, with new harness replicas starting first and old ones draining rather than dying.
The runtime speaks standard interfaces. It supports A2A version 1.0, so incoming agent tasks become real sessions, and MCP in both directions: the harness connects to platform MCP servers and exposes its own operations as MCP tools. It also works with an API compatible with the Claude Managed Agents quick-start flow.
PrimusClaw is MIT licensed, except for the sandbox control plane, a fork of the Kubernetes SIG Apps agent-sandbox project under Apache 2.0. For teams already running long agent jobs on their own hardware, such as the coding agents built on Devin, it offers a self-hosted alternative to managed services — with the trade-off that compliance and service levels become the operator’s job.
Source: AMD ROCm Blog
