DigitalOcean Bets on a Managed Runtime for AI Agents

Running an AI agent in production usually means someone on the team becomes an infrastructure engineer by accident: standing up a sandboxed runtime, wiring up tool access, picking a model, and building a way to pause and resume long-running sessions without losing state. DigitalOcean's answer, launched into public preview this week, is to sell that whole stack as one product. Managed Agents lets a developer run Claude Code, Codex CLI, OpenCode, Hermes or a custom LangGraph agent inside an isolated cloud runtime, hand it governed access to more than 16,000 tools, and pay only for the seconds the agent is actually thinking.

Table of contents

• What Managed Agents actually bundles
• The three pieces underneath it
• Why pause and resume is the interesting detail
• How the pricing works
• The gap it's aimed at
• What it doesn't solve
• Frequently asked questions

What Managed Agents actually bundles

DigitalOcean's own launch post describes the pitch directly: developers pick an agent harness they already use, connect it to tools, and let it do meaningful work in seconds, without building or maintaining the infrastructure themselves. That is a real category of unglamorous engineering work. Before this kind of product existed, a team running Claude Code or Codex CLI at scale had to handle sandboxing, tool permissions, session state and model routing as four separate problems, often on four different vendors.

Support for Claude Code, Codex CLI and OpenCode, plus general-purpose agents like Hermes and anything built with LangGraph, comes unmodified, according to DigitalOcean's product page. Teams can also package a custom agent as a standard Open Container Initiative image and reuse it as a template, which matters for anyone who has already built an internal agent and does not want to rewrite it to fit someone else's platform.

The three pieces underneath it

Investing.com's coverage breaks the product into three named components. Harness Runtime combines a microVM with built-in tools and coding sandbox capabilities, giving each session an isolated machine to run in. Action Gateway provides governed access to those 16,000-plus tools from more than 500 providers through a single managed endpoint, rather than each agent needing its own individually configured integrations. Inference Engine connects agents to DigitalOcean-hosted open-weight models such as Nemotron 3 Ultra, Kimi K3 and GLM 5.3, to proprietary models including Claude Fable 5.1 and GPT 6 Astra, or to a customer's own model endpoint, with an Inference Router that picks a model based on intent, cost, latency or quality.

Put together, the pitch is that a team no longer chooses between a coding harness, a tool-access layer and a model provider as three separate vendor relationships. DigitalOcean is betting that bundling those three decisions under one bill is worth more to a mid-sized engineering team than picking the single best option in each category separately.

Why pause and resume is the interesting detail

The specific engineering claim worth noting is speed of state recovery. DigitalOcean's product documentation says a session that finishes its assigned work pauses and snapshots its state, and can be reattached later, from the same device or a different one, picking up exactly where the agent left off. Investing.com's figure for that resume time is 305 milliseconds; DigitalOcean's own blog post says roughly 300 milliseconds.

That number matters more than it sounds. Long-running agent tasks, the kind that take minutes or hours rather than seconds, are expensive to keep paying for while they sit idle waiting on a human or another system. A near-instant pause and resume means a team can let an agent go quiet between steps without losing the accumulated context of the session, and without paying compute costs during the gap.

How the pricing works

Investing.com reports CPU is billed at $0.044 per vCPU-hour and memory at $0.0095 per GB-hour, active-usage pricing rather than a flat monthly rate. DigitalOcean says it charges no CPU fees while an agent is waiting on a model response or a tool result, which is the more meaningful commitment: a slow tool call or a model that takes a while to respond does not silently run up the bill the way idle compute in a traditional server would.

New DigitalOcean users get a $5 credit to launch a first agent session, according to the company's press release, a low-friction way to try the product before committing to production use.

The gap it's aimed at

DigitalOcean's own framing draws an explicit comparison to an earlier cloud shift. Its blog post quotes the line that EC2 was the front door to the first generation of cloud computing, where customers rented a virtual machine and assembled everything else around it. Managed Agents is DigitalOcean's bet that the equivalent front door for the agent era is not a virtual machine at all, but a managed runtime built specifically for AI agents doing multi-step, tool-using work.

Amplitude co-founder and CEO Spenser Skates, quoted in the launch materials, said the company plans to use DigitalOcean's Harness Runtime and Action Gateway for fast startup and scaling, intelligent tool discovery and secure tool access, freeing engineering time to focus on the product itself rather than the plumbing underneath it. Qencode, a media processing company, is cited as an early customer running a support-triage agent that reviews incoming requests across communication channels and manages Jira tickets through the platform.

What it doesn't solve

A managed runtime and a governed tool gateway address the operational side of running agents, not the security questions that come with giving any agent broad tool access. Our earlier report on the Plugin4Shell vulnerability covered how plugin systems in coding agents can be compromised even when a marketplace has approved the exact version installed. Centralizing tool access through a single gateway, as Action Gateway does, could make that kind of auditing easier in principle, but DigitalOcean's launch materials do not describe specific protections against a compromised plugin or a malicious tool reaching an agent through the gateway.

It is also, for now, a public preview rather than a generally available product, which typically means terms, pricing and feature scope can still shift before wider rollout. Teams evaluating it for anything beyond a pilot project should read the preview terms DigitalOcean requires customers to opt into before treating any of the current numbers as fixed.

Frequently asked questions