OpenClaw MLOps Proposal

OpenClaw contributors are discussing a cloud rollout plan that would manage autonomous agents more like machine learning systems than one-off virtual machines. (github.com, techcommunity.microsoft.com) OpenClaw is open-source software for running a self-hosted, always-on agent on infrastructure a company controls, instead of inside a vendor’s chat app. Microsoft published an Azure deployment guide on March 22, 2026, showing OpenClaw on Azure Linux virtual machines with Network Security Group rules and Azure Bastion access. (github.com, techcommunity.microsoft.com, docs.openclaw.ai) That Azure guide uses command-line provisioning for a resource group, virtual network, subnets, a Bastion host, and a virtual machine, with a sample size of Standard_B2as_v2 and a 64 gigabyte operating system disk. It is a secure single-deployment pattern, not a fleet-management system for large numbers of agents. (docs.openclaw.ai) Machine learning operations, or MLOps, is the practice of packaging, deploying, updating, and observing machine learning systems with repeatable controls. Amazon’s machine learning blog describes MLOps platforms as built for reproducibility, robustness, and end-to-end lifecycle management through code repositories and automated deployment. (aws.amazon.com) Applied to agents, that model would treat an agent image like a sealed machine snapshot that can be copied fast, replaced cleanly, and upgraded without losing its memory or configuration. OpenClaw already has users pushing toward that direction on other clouds: an AWS Terraform module stores agent data on Elastic File System so instances can be replaced while state persists. (github.com) The proposal surfaced as OpenClaw’s GitHub project showed 356,000 stars and 72,000 forks on April 13, 2026, with the codebase updating daily. That scale has pulled the project beyond hobby deployments and into the kind of operational questions large companies ask about patching, rollback, and standard builds. (github.com) OpenClaw’s own documentation already frames the software as a long-running runtime that can connect to Microsoft Teams, Slack, Telegram, WhatsApp, and other channels while using model providers such as Azure OpenAI, OpenAI, Anthropic Claude, Google Gemini, and GitHub Copilot. A lifecycle layer would sit under those integrations and standardize how the agent is provisioned and maintained. (techcommunity.microsoft.com, clawdocs.org) The open question is portability. The Azure material from Microsoft focuses on Azure command-line setup, while the existing third-party Terraform module targets Amazon Web Services with Cognito, Application Load Balancer, CloudWatch, and Elastic File System. (techcommunity.microsoft.com, github.com) That leaves the current discussion centered on whether OpenClaw should add a first-class lifecycle layer inside the project or keep relying on separate cloud-specific deployment guides and community modules. Either way, the immediate shift is from “how do I start one agent” to “how do I run and update many of them.” (docs.openclaw.ai, github.com, github.com)