Oxide Computer Company Offers Guidance for Running Self-Hosted AI Agents in Virtual Machines

Oxide Computer Company has released new reference architectures and examples for deploying self-hosted AI agents within virtual machines on its Oxide Cloud Computer. The guidance emphasizes that while AI agents can be run similarly to other workloads in VMs, their unique characteristics — such as the blend of GPU-based inference and CPU-based tool execution — necessitate a more nuanced approach. Specifically, Oxide recommends running the AI agent's orchestrator as a persistent virtual machine, which then dynamically provisions ephemeral virtual machines for individual AI agent sessions. This model is demonstrated using popular environments like Claude Code and Amp self-hosted orbs. This development is significant for cloud and DevOps practitioners because it provides a clear, actionable strategy for managing the lifecycle and resource allocation of AI agents. The ephemeral VM approach for agent sessions offers several advantages: enhanced security through isolation, efficient resource utilization by spinning up resources only when needed, and simplified cleanup. It also allows for greater control over the execution environment, which is critical for debugging, auditing, and ensuring compliance. For organizations that are already heavily invested in virtualized infrastructure, this guidance offers a familiar and robust path to integrate AI agents into their existing operational paradigms. The broader trend in cloud and DevOps is a growing focus on specialized infrastructure for AI workloads. While general-purpose cloud resources can host AI, the demand for optimized environments that cater to the specific needs of AI agents – such as low-latency access to GPUs, efficient data pipelines, and robust orchestration – is increasing. This move by Oxide aligns with the industry's shift towards more prescriptive architectures for AI, recognizing that agents are not just another application but a distinct workload type requiring dedicated consideration for deployment and management. Other platforms like Microsoft Agent Framework, Google ADK, and various open-source frameworks like LangGraph and CrewAI are also emerging to address the need for structured agent development and orchestration, highlighting the industry-wide recognition of these challenges. In practice, this means that practitioners should evaluate their current infrastructure capabilities and consider how a VM-centric approach could benefit their AI agent deployments. Key considerations include the overhead of managing additional VMs, the efficiency of dynamic provisioning, and the integration with existing orchestration and monitoring tools. Teams should focus on automating the provisioning and de-provisioning of these ephemeral agent VMs to minimize operational burden. Furthermore, understanding the interplay between the agent's inferencing (GPU) and tool execution (CPU) requirements will be crucial for optimizing VM sizing and configuration. This guidance from Oxide serves as a valuable blueprint for building scalable, secure, and manageable AI agent systems, pushing practitioners to think critically about the underlying compute environment for their agentic applications.