DigitalOcean Launches Managed Agents, Removing the Infrastructure Barrier to Running AI at Work

30-Second Summary

DigitalOcean launched Managed Agents in public preview on 22 September 2026. The product gives every AI agent its own isolated execution environment, connects it to 16,000 plus business tools through a single managed MCP endpoint, and lets it choose from 75 plus AI models. Billing runs only when the agent is active. Supported harnesses include Claude Code, Codex CLI, OpenCode, Hermes, and LangGraph. This is DigitalOcean's most direct move yet into the AI infrastructure market, targeting developers and small-to-mid-size businesses that want to run agents in production without building and operating the surrounding platform themselves.

At a Glance

• Topic: Agent Systems
• Company: DigitalOcean
• Date: 22 September 2026
• Announcement: Managed Agents public preview, including isolated harness runtimes, Action Gateway for governed MCP tool access, and Serverless Inference across 75 plus models
• What Changed: Previously, running an AI agent required assembling execution infrastructure, tool integrations, model access, and billing tracking separately. DigitalOcean consolidates all four under one cloud account and one bill.
• Why It Matters: The infrastructure overhead has been the main reason AI agents remain in pilots rather than production at smaller businesses. Managed Agents removes that barrier with a pay-per-use model and a security architecture teams can explain.
• Who Should Care: Developers and technical leads at companies with 10 to 200 employees who want to deploy AI agents against real business tools without a dedicated platform engineering team

Key Facts

• Launch date: 22 September 2026, public preview
• Harnesses supported: Claude Code, Codex CLI, OpenCode, Hermes, LangGraph
• Tool access: 16,000 plus tools across 500 plus providers via Action Gateway (single managed MCP endpoint)
• Providers included: GitHub, HubSpot, Stripe, Snowflake, PagerDuty, Box, Supabase, Exa, web search, browser automation, code execution, and custom MCP servers
• Model choice: 75 plus open and proprietary models via Serverless Inference
• Billing: CPU billed only when the agent is active. Paused sessions retain state with no CPU or memory charges.
• Access: Public preview to all DigitalOcean customers who opt in to preview terms. New users receive a $5 credit.
• Interface: Cloud Console or doctl CLI. Tool access configured through Action Gateway using the DigitalOcean SDK or any MCP-compatible client.

What Happened

DigitalOcean announced the public preview of Managed Agents on 22 September 2026, describing it as "agent execution, tool access, and inference together on one cloud." The announcement positions the product as a complete stack rather than a component. Developers create agent sessions in the Cloud Console or through the doctl CLI, selecting a harness (Claude Code, Codex CLI, OpenCode, Hermes, or LangGraph) and connecting tools through Action Gateway.

Action Gateway is the mechanism that differentiates Managed Agents from earlier DigitalOcean AI products. It acts as a single managed MCP endpoint, handling authentication to connected services on behalf of the agent so that developers do not need to share credentials directly. The gateway covers 500 plus providers across 16,000 plus individual tools. Businesses can also connect their own MCP servers through the same endpoint.

Each agent session runs in its own isolated microVM, a lightweight virtualisation unit that provides resource and process isolation between concurrent sessions. When a session is idle, it pauses and retains its state, and CPU billing stops. When work resumes, billing restarts. This pause-and-resume approach is designed specifically for agentic workloads that may run in bursts rather than continuously.

Serverless Inference is included in the same platform, giving agents access to 75 plus models from open and proprietary providers. The billing structure runs from the same DigitalOcean account, consolidating what previously required separate API keys and invoices across multiple providers.

Why It Matters

The infrastructure problem was always the blocker, not the model. Teams that could not ship AI agents to production were not stuck on capability. They were stuck on where to run the agent, how to connect it to internal tools safely, how to manage credentials, and how to prevent spiralling cloud costs. Managed Agents addresses all four.

Governed tool access is the compliance story many businesses have been missing. Action Gateway gives organisations a single point to review, scope, and revoke what any agent can do. For regulated industries, or any company with client data in its systems, this matters more than raw capability. The alternative, running agents with unrestricted API access to business tools, is not a governance architecture most organisations can defend.

Claude Code as a first-class harness is significant. DigitalOcean has named Claude Code alongside Codex CLI and OpenCode as supported harnesses. This is not incidental support. It means businesses running Claude Enterprise or Claude Team can extend those deployments into DigitalOcean-managed agent infrastructure with a clear integration path, without managing separate cloud environments.

Pay-when-active billing changes the economics of agent pilots. One of the most common objections to running more agents is cost uncertainty. Managed Agents addresses this by pausing billing when the agent is idle. For businesses running agents on recurring tasks rather than continuous pipelines, the effective cost is substantially lower than provisioning always-on compute.

DigitalOcean's market is the business that most needs this. DigitalOcean built its business serving developers at companies too small for AWS enterprise support and too technical for no-code tools. Managed Agents is aimed directly at that segment. A 40-person business wanting to run a HubSpot-connected sales agent on Claude Code no longer needs a cloud architect to make it happen.

The David and Goliath View

The announcement that will define AI in business over the next 12 months is not a new model. It is the infrastructure layer that makes running agents simple enough to be a normal business decision rather than a technical project. DigitalOcean has made a credible move toward that layer with Managed Agents. The combination of isolated execution, governed MCP tool access, and pay-per-use billing is exactly the architecture an operator needs to say yes to AI agents in production rather than keeping them in a controlled demo.

Action Gateway is the part that deserves the most attention. MCP (Model Context Protocol) defined how AI systems connect to tools. Action Gateway operationalises that standard for businesses that do not want to build their own integration infrastructure. When a team can say "our agent connects to HubSpot and Stripe through a single governed endpoint that lives in our DigitalOcean account," they have an AI architecture story that holds up to scrutiny from finance, legal, and security.

For businesses already using Claude, the Managed Agents announcement closes a gap. You could run Claude in a chat interface or through the API, but deploying Claude Code as an agent with persistent state, governed tool access, and per-use billing required assembling those pieces yourself. That assembly work is now done. The path from a Claude Enterprise subscription to a production agent doing real work in your business systems is shorter than it has ever been.

Where This Fits in the AI Stack

Managed Agents sits at the execution and governance layer, the part of the AI stack that connects models to real business tools and ensures they only do what they are supposed to do. It addresses the gap between a capable model and a deployed agent doing useful work.

• Models: DigitalOcean's Serverless Inference (75 plus options including Claude)
• Harnesses: Claude Code, Codex CLI, OpenCode, Hermes, LangGraph (agent logic layer)
• Tool access: Action Gateway (single MCP endpoint, 16,000 plus tools)
• Execution: Managed microVM runtimes (isolated, pause-and-resume)
• Billing: One DigitalOcean account (execution, inference, tools consolidated)

Questions Operators Are Asking

Is this production-ready or still experimental? It is in public preview, which means DigitalOcean considers it stable enough for real workloads but is still gathering feedback. The harnesses it supports (Claude Code, LangGraph) are production tools. For non-critical workloads, this is deployable today. For systems touching financial or client data, a careful pilot with scope-limited tool permissions is the prudent path.

Do I need to be a DigitalOcean customer already? No. Any developer can opt in to the public preview. New accounts receive a $5 credit toward the first session.

How does Action Gateway handle tool credentials? Action Gateway manages authentication to connected providers through the DigitalOcean platform. The agent accesses tools through the gateway endpoint rather than receiving credentials directly. This means credentials are not exposed to the agent runtime and can be scoped, audited, and revoked through the DigitalOcean console.

We already use Claude Enterprise. Does this replace that or complement it? It complements it. Claude Enterprise gives your team access to Claude through Anthropic's platform. Managed Agents provides the execution infrastructure and tool connectivity to run Claude Code as an agent in production workflows. Both can be active simultaneously.

What happens to an agent session when it pauses? State is retained. When the next task arrives, the session resumes from where it left off. CPU and memory billing stops during the pause period, which means you are not paying for idle time.

Citable Summary

DigitalOcean launched Managed Agents in public preview on 22 September 2026. The product provides isolated microVM execution environments for AI agents, a single managed MCP endpoint called Action Gateway with access to 16,000 plus tools across 500 plus providers, and Serverless Inference across 75 plus models. Supported harnesses include Claude Code, Codex CLI, OpenCode, Hermes, and LangGraph. Billing pauses when agents are idle. The product targets developers and businesses that want to run AI agents in production without building and managing agent infrastructure separately. It is available to all DigitalOcean customers who opt in to the public preview terms.