DigitalOcean Launches MicroVMs in Private Preview for AI Agent Workloads

Why agents strain traditional infrastructure

According to DigitalOcean, an AI agent uses compute differently than a traditional application. In a single session it writes and runs code, calls a model, and waits on a tool or a person before continuing. That means it needs real compute in short, intensive bursts, plus a hard boundary around the code it has just generated so that one agent's mistake cannot spread to the host or to other customers.

The company argues that existing options fall short. An always-on server for a workload that is busy only a few minutes an hour means paying for a lot of idle time. Containers start quickly and pack densely, but they share a kernel with their neighbors, which DigitalOcean describes as more trust than you want to place in code an agent just wrote. Teams can close the gap themselves with pausing, snapshots, fast startup and per-tenant isolation, but that requires running their own virtualization platform.

How MicroVMs work

A MicroVM is a lightweight virtual machine that runs a container image in its own VM, with its own kernel, isolated from other workloads on the host. Users bring a public or private image from DigitalOcean Container Registry, choose a size, and reach the running workload over an authenticated HTTPS endpoint. The MicroVMs run on open-source Firecracker technology, so each VM has a hardware-virtualized boundary rather than a shared kernel.

The key difference from a traditional VM is behavior when work stops. A MicroVM pauses automatically after a timeout the user sets and resumes on the next request with its memory, files and running processes as they were. It consumes compute only while it is running.

Checkpoints address the setup cost on every start. A user can boot a MicroVM once, let it load its runtime, dependencies and models, then checkpoint that warmed-up state and start new MicroVMs from it instead of rebuilding the environment each time. DigitalOcean says it runs the fleet, the isolation and the lifecycle underneath.

Intended use cases

DigitalOcean positions MicroVMs first for agent infrastructure, where each session needs its own isolated environment that starts on demand, pauses while waiting and resumes with state intact. It lists these fits:

• Coding and eval workspaces: a hardware-isolated MicroVM per task, disposed of when the agent is done.
• Long-running agent workflows: workflows that run for hours, pause while waiting on a model or a person, and resume where they left off.
• Parallel reinforcement learning environments: many identical environments side by side, reset sub-second for the next episode.
• Ephemeral data per agent: a throwaway Postgres or scratch store whose data survives a pause and is cleared when the session ends.
• Per-tenant tools and MCP servers: each customer's tools in their own VM.
• Untrusted or generated code: plugin sandboxes, security analysis and evaluation jobs.
• Short-lived jobs: CI/CD runners, pull request preview environments and ETL steps.

Relationship to Managed Agents and availability

DigitalOcean Managed Agents, available in public preview, runs every agent session inside a MicroVM. According to the company, its execution layer, Harness Runtime, uses MicroVMs to isolate agent-generated code, pause a session the moment it goes idle, and resume it in under a second with files and state intact. Teams at OpenHands, Qencode and Amplitude are already building on Managed Agents.

The two products serve different needs. Managed Agents is the hosted platform where a customer brings an agent and DigitalOcean operates the runtime, tool access and scaling. MicroVMs are the underlying layer, aimed at teams that prefer to build and control their own execution layer. DigitalOcean suggests using them to build an agent platform, a sandbox product or a code-execution service.

MicroVMs are currently available only through an invite-only private preview, and interested teams can request access. The source does not give pricing or a general availability date.