On November 6, 2026, AMD introduced the open PrimusClaw runtime for AI tasks on private GPU clusters. It reports more than 40,000 internal sessions. These are company operational figures, not an independent comparison.

Keeping sessions alive beyond a worker

The runtime separates task state from workers and command execution. Its stated recovery mechanism uses checkpoints; having that mechanism does not replace evaluation on an operator's own workload.

The repository explains the implementation. An API accepts sessions; Brain manages model requests and tools; Hands performs file and command operations in a separate environment. PostgreSQL stores state, NATS handles queues, and S3-compatible storage synchronizes workspace files. Production deployment needs Kubernetes and Primus-SaFE for authentication and resource placement. The main code is MIT-licensed, while the derived sandbox component has a separate Apache 2.0 license.

That distinguishes available software from a ready-made service. Operators maintain their own installation and must meet the terms of external models and tools. An open license for the orchestrator does not make compute or model requests free.

A boundary for untrusted code

The security policy assumes sandbox code can be hostile. It requires authentication, restricted outbound access and suitable isolation; ordinary containers share the host kernel. It also warns that sensitive-field redaction cannot guarantee removal of every possible secret. These are requirements and known limits, rather than a certificate of invulnerability.

Editorial assessment: how to evaluate usefulness

A sensible first evaluation would be one authorized optimization task on a test copy of code. Define success beforehand: whether the program remains correct, which resources it uses, and whether the resulting change can be reproduced. A long session with many tool calls is not evidence of a useful optimization by itself.

Recovery is the next question. An operator could interrupt a test worker and examine whether the agent continues from the right state without repeating an external action already completed. A conversation checkpoint must remain consistent with the actual files. This is a proposed evaluation criterion, not a test SIExpo has performed.

Budgeting also matters on a shared cluster. Each task should have its own time, accelerator and external-request limits. Otherwise, making agents easier to start may simply accelerate overspending. The useful organizational outcome is a verified change with a known cost and accountable owner, rather than a large count of concurrent sessions.

PrimusClaw supplies a concrete open foundation for that management. SIExpo has not deployed the project and does not claim independently measured performance, security or savings.

Editorial verification: 2026-11-08T17:56:23.508205+00:00.