For a long time, we've treated AI agents like chatbots with a few plugins. But the new wave of 'agentic' tools—like Claude Code and Cursor—isn't just predicting the next token; they're actually doing work. To do that safely, they need more than a prompt window. They need a home. Enter the specialized Virtual Machine (VM).
Beyond the Sandbox
Standard containers are great for apps, but AI agents need 'agent-scale' infrastructure. We're seeing a shift toward micro-VMs, specifically using technologies like Firecracker. These aren't your clunky old server VMs; they are lean, isolated environments that can boot in milliseconds.
Companies like Freestyle are pushing this to the limit, offering full Linux VMs with root access that provision in about 65ms. This allows an agent to install a package, spin up a server, or run a compiler without risking the host system's security. As one developer noted, the environment has become just as important as the model itself.
Statefulness and 'Memory in Git'
One of the biggest hurdles for mobile agents is state. If an agent is working on your codebase across a browser, a desktop app, and a mobile device, it can't just lose its place.
Intriguingly, some platforms are implementing 'memory in git.' By treating the VM's state or the agent's progress as a version-controlled history, agents can maintain a persistent, stateful environment. This turns the VM from a temporary scratchpad into a living workspace that evolves as the agent solves complex bugs or builds new features.
The Future of Agentic Compute
We are moving toward a world where every AI agent carries its own portable OS. Whether it's through custom agent definitions in .claude/agents/ or integrated cloud VMs, the goal is a seamless transition between the LLM's reasoning and the OS's execution.
Sources
Media
