DigitalOcean opens MicroVMs public preview for AI agent workloads

What a MicroVM is and how it works

According to DigitalOcean, coding-agent platforms, sandbox products and code-execution services all need isolated machines that start quickly, keep their state between bursts of work and use no compute while idle. Building that in-house means leasing capacity and turning raw servers into on-demand compute, including fleet management, snapshots, pause and resume, per-tenant isolation, placement and recovery.

A DigitalOcean MicroVM is an isolated virtual machine built on Firecracker. Users create one from a container image that packages the agent, its toolchain and its dependencies. The MicroVM boots with its own kernel and a hardware-virtualized boundary rather than a shared one. DigitalOcean says the design is meant to contain any bad steps by an agent to that single MicroVM, isolated from the host and other customers. Sessions are reached over an authenticated HTTPS endpoint, so only requests carrying a valid DigitalOcean API token can connect.

Pause, resume and checkpoints

The company points out that a coding agent spends much of a session waiting for a model response, for code to run or for a person to review its work. A MicroVM pauses after an idle period that the user sets, and compute is not charged while it is paused. When the next request arrives, DigitalOcean says the MicroVM resumes in milliseconds with its memory, files and running processes unchanged.

Startup is also in milliseconds, according to the company. A burst of new sessions simply starts more MicroVMs on demand, up to the team's MicroVM limit, so capacity does not have to be provisioned in advance.

Users can also set up an environment once, checkpoint it and start new sessions from that checkpoint, so each session begins pre-warmed instead of being rebuilt from scratch.

• Automatic recovery: if a workload crashes after starting, the MicroVM restarts it. A workload stuck in a failure loop stops retrying, which prevents a broken image from restarting endlessly.
• Private networking: MicroVMs can run inside a customer's VPC to reach databases and internal services without exposing them publicly. DigitalOcean can create a VPC if the customer has none.

Link to Managed Agents and pricing

MicroVMs also power DigitalOcean Managed Agents, which is available in public preview. MicroVMs provide the virtual machine and its lifecycle: creating it from an image, pausing and resuming, checkpointing, running commands and access through an endpoint. Managed Agents sits above that, and its DigitalOcean Harness Runtime adds an agent control plane covering file reading, writing, upload and download, running and upgrading the agent harness, preserving session state across runs and brokering governed access to tools.

Compute is billed per second only while a MicroVM runs, at $0.022 per vCPU-hour and $0.0085 per GiB-hour. Storage costs $0.05 per GiB-month and keeps the disk and saved state available, so a paused MicroVM incurs only storage charges. Outbound data transfer on public networking is $0.01 per GiB.

DigitalOcean gives an example: a MicroVM with 2 vCPU and 4 GiB of memory running a full hour costs $0.078 in compute. A coding-agent session on that configuration that runs 15 minutes of the hour and is paused for the rest would cost $0.0195 in compute, plus storage.

Getting started and what it means

Beyond coding agents, DigitalOcean lists sandboxes for untrusted code, evaluation and reinforcement learning environments, CI/CD runners and pull request preview environments as possible uses. A MicroVM can be created from the Control Panel, the doctl command line tool or the API. The default HTTP port is 8080, and the --http-port flag can match a different container port. After creation, users can pause, resume, checkpoint and run commands inside it.

For teams building agent products, the offer is a managed alternative to running their own isolation and fleet layer, with pricing tied to active time rather than session length. The service is still a public preview, and the source does not give regional availability details beyond the region option in the create command.