DigitalOcean Managed Agents: microVM sessions and a gateway to 16,000 tools
Victoria Shinder2 min readDigitalOcean has launched DigitalOcean Managed Agents in public preview, InfoQ reports, a managed infrastructure layer for running AI agents with isolated microVM runtimes, governed tool access and serverless inference.
DigitalOcean's argument is that agents differ from ordinary cloud applications: developers who run them on conventional VMs end up building plumbing to preserve an agent's context, keep its artifacts accessible beyond the agent that created them, coordinate parallel work and secure its access to tools, and they keep spare capacity so agents start quickly. The product combines two services aimed at that work.
The Harness Runtime, built on a lightweight microVM, gives agents persistent, isolated compute across supported harnesses, including coding agents such as Claude Code, Codex CLI and OpenCode, general-purpose agents such as Hermes, and custom agents built with LangGraph. It keeps conversation history and working state across sessions, which can be paused, resumed or forked, and pauses agents automatically when idle, meaning no ongoing model or tool calls. Sessions can reach internal services without exposing them publicly and can run in parallel across repositories and tasks.
The Action Gateway is the agent's interface to external tools through one MCP endpoint, with more than 16,000 tools available, from web search and browser automation to DigitalOcean's own infrastructure APIs and connectors for GitHub, HubSpot and Stripe. It manages permissions centrally, can require human approval for sensitive operations, and handles rate limiting, retries, backoff and timeouts. InfoQ notes the overlap with Docker's Cloud Sandboxes, which also offer microVM isolation and persistence but focus on moving sandboxes between local and cloud machines, while DigitalOcean's product leans towards production orchestration and tools.
Why it matters
Idle agents are mostly a cost problem: a coding agent waiting for a review still holds a machine. Pausing on idle and forking sessions are the kind of features that decide whether running many agents in parallel is affordable, and they are now offered as a managed service rather than something each team builds.