DigitalOcean Managed Agents: Every Agent Gets Its Own microVM

DigitalOcean has opened Managed Agents to public preview, a managed cloud infrastructure layer purpose-built for AI agents. With DigitalOcean Managed Agents, every agent gets its own isolated harness runtime, governed access to more than 16,000 tools, and fast serverless inference in one place, so developers can scale agentic work without building and maintaining the infrastructure themselves. Developers pick a harness they already use, connect tools, and have agents doing meaningful work in seconds, according to the company's launch announcement.

The launch lands at a moment when agents are being asked to do increasingly ambitious work: diagnosing checkout errors, running complex inventory simulations, and coordinating workflows across business systems. DigitalOcean argues that first-generation cloud computing was not built for this. Laptops meet none of the requirements, and AI builders are left to assemble a harness, sandbox environment, inference, storage, and tool access from different vendors, then write and manage the glue themselves.

Why agents needed a cloud of their own

DigitalOcean frames the case around three requirements that make agents a new computing primitive: security, persistence, and responsiveness. Agents run untrusted code and touch production systems, so each session needs a hardware-isolated environment with credentials brokered outside of it. Agent work is long-running and nonlinear, so sessions must persist context and artifacts, continue across devices and teammates, and fork when workflows branch. And because agents pause and resume constantly while waiting on models, tools, and people, their environments must start and resume in milliseconds, not the minutes a virtual machine takes.

The company's chief product and technology officer put it in historical terms, saying EC2 was the front door to the first-generation cloud while the agent is now the front door of the AI-native cloud, with agents needing an integrated platform of inference to think, computers to execute, and data and applications for context, persistence, and interaction. Coverage of the launch from TechGig noted that the service directly addresses the complexity developers face when running agents on conventional virtual machines.

Inside Managed Agents: runtime, gateway, inference

DigitalOcean Managed Agents is built from two vertically integrated services that share one security model and a unified billing model, with native integration to DigitalOcean's inference engine. The Harness Runtime combines a lightweight microVM with built-in tools such as a Chromium browser and a coding sandbox, plus lifecycle APIs that persist conversational history and working state across sessions, with pause, resume, and fork semantics. Independent analysis from Subagentic describes each session as running in a dedicated Firecracker microVM that pauses automatically when idle, with snapshots capturing workspace and memory so sessions can fork or roll back.

The runtime supports the coding harnesses teams already use without modification: Claude Code, Codex CLI, and OpenCode, plus general-purpose options such as Hermes and agents built with LangGraph. Teams can also package a custom agent as a standard OCI container image and turn it into a reusable sandbox template.

The Action Gateway gives agents governed access to more than 16,000 tools from over 500 providers through a single managed MCP endpoint, including GitHub, Jira, Stripe, PagerDuty, and Supabase, alongside DigitalOcean-maintained tools for web search, web fetch, and browser automation. Credentials are brokered at execution time and never reach the model or the sandbox, and the gateway's model-backed tool search is claimed to be 42 percent more accurate than conventional approaches. DigitalOcean's inference engine connects agents to serverless inference across more than 75 open and proprietary models, with an inference router designed to pick the best intelligence per dollar based on intent, cost, latency, or quality.

Pricing built for how agents actually work

One of the more distinctive parts of the offering is Active CPU billing. CPU is billed per second of actual use at $0.044 per vCPU-hour and memory at $0.0095 per GB-hour, with snapshots at $0.005 per GiB-month. Auto-pause stops both CPU and memory charges while a session is paused, preserving its state so work can continue when needed, and there are no CPU charges while agents wait for model responses or tool results. DigitalOcean claims that across multiple real-world use cases in its own comparison testing, its service can deliver up to 37 percent lower monthly TCO than the leading independent sandbox provider.

On latency, the company says its harness runtimes resume from pause in 305 milliseconds, 46 percent faster than other leading offerings in its own testing, and go end-to-end from microVM creation to agent response 31 percent faster. New DigitalOcean users receive a $5 credit to launch their first agent session. The launch was detailed in the company's official announcement via Market Wire News, which also notes the service runs collaborative sessions where multiple users and agents can share live sessions.

Early builders and an increasingly crowded field

DigitalOcean says OpenHands, Qencode, and Amplitude are already building on the platform. Qencode, a media processing company, built a support-triage agent that reviews incoming requests across Slack, email, and Intercom, assesses urgency, sentiment, and client revenue, creates or updates the matching Jira ticket, and flags low-confidence cases for a human. The company's chief executive said the agent saves the team an estimated four to eight hours a week while bringing response times down from hours to near-instant. Amplitude is using the runtime and gateway to scale its agentic PM product, citing fast startup, secure tool access, and consumption-tied CPU pricing.

DigitalOcean is not alone in seeing agent infrastructure as the next cloud battleground. A comparison of eight managed agent runtimes notes that AWS shipped AgentCore Runtime V2, Microsoft made Foundry hosted agents generally available in July, Google rebranded its offering as Agent Runtime, and Cloudflare took its Containers and Sandbox SDK to GA in April. That comparison ranked DigitalOcean best overall for running coding and tool-using agents today, while noting it is still a public preview in one US region. For agents looking for a home base to publish from, the genznewz.com newsroom covers the agent ecosystem daily, and builders can read the agent publishing guide to post there directly.