DigitalOcean has introduced DigitalOcean Managed Agents in public preview, positioning the service as a managed cloud infrastructure layer for AI agents. The new offering combines isolated microVM runtimes, controlled tool access, and serverless AI inference to reduce the amount of plumbing developers need to build themselves when running agentic workloads in production.
What DigitalOcean Managed Agents is meant to solve
According to DigitalOcean, agentic workflows are materially different from traditional cloud applications. Developers running agents on conventional virtual machines often have to assemble the surrounding runtime themselves, including context persistence, artifact storage, parallel coordination, and security-hardened access to tools.
The company also points to operational overhead that becomes especially painful at scale. Teams need spare VM capacity so agents can start quickly, and they have to provision and configure compute resources on demand. DigitalOcean says Managed Agents is designed to take on that infrastructure burden so developers can focus on the agent logic itself.
Two services: Harness Runtime and Action Gateway
DigitalOcean Managed Agents is built around two connected services: the Harness Runtime and the Action Gateway. Together, they are intended to let developers scale agent work while DigitalOcean manages execution, persistence, tool access, and the infrastructure layer underneath.
Harness Runtime for isolated, persistent agent sessions
The Harness Runtime is based on a lightweight microVM and provides persistent, isolated compute environments for agents. DigitalOcean says it supports a range of harnesses, including coding agents such as Claude Code, Codex CLI, and OpenCode; general-purpose agents such as Hermes; and custom agents built with LangGraph.
One of the runtime’s main features is the ability to persist conversational history and working state across sessions. Those sessions can be paused, resumed, or forked, which is important for workflows that do not finish in a single uninterrupted run. The company also says sessions can automatically pause agents when they are idle, meaning when there are no ongoing LLM or tool calls.
Each session runs on security-hardened compute and storage resources. DigitalOcean says developers can connect to internal services without exposing them publicly, which is a key requirement for many production environments. Sessions can also be launched in parallel across repositories and tasks, supporting divide-and-conquer, collaboration, and map/reduce-style workflows.
Action Gateway for governed tool access
The Action Gateway serves as the agent’s interface to external tools and services through a unified MCP endpoint. DigitalOcean says more than 16,000 tools are currently available through the gateway, including Web Search, Web Fetch, Browser Automation, DigitalOcean infrastructure management APIs, and connectors for platforms such as GitHub, HubSpot, Stripe, and others.
Security and operational control are central to the design. The gateway includes centralized permission management so teams can define which tools and actions are available to each agent. It also requires human approval for sensitive operations, adding a review step when an action could have significant effects.
To support multiple agents accessing external systems efficiently, the gateway also provides rate limiting, retries, backoff, and timeouts. Those controls matter in real-world deployments, where agent traffic can be bursty and integrations may fail or slow down unpredictably.
Why this matters for production AI agents
The public preview reflects a broader shift in how AI agents are being deployed. Early demonstrations often focus on a single agent completing a task in a sandbox, but production use cases tend to involve longer-lived workflows, state that must survive restarts, and access to internal or third-party systems that cannot be exposed casually.
That is the space DigitalOcean is targeting. By packaging isolation, persistence, tool governance, and orchestration support into one service, the company is aiming at teams that want to run agents without building a custom control plane around them. The emphasis on managed infrastructure is especially relevant for teams that expect to run many agents in parallel or keep them active across multiple sessions.
The mention of serverless AI inference in the product description also suggests that the platform is being shaped for workloads that need to invoke models as part of a wider workflow, rather than treating the model call as the only part of the system. In practice, that means the surrounding runtime, state, and tooling matter almost as much as the model itself.
Industry reaction highlights the cost and operations angle
The InfoQ report cites reactions on X from people working in the AI infrastructure space. Former Meta AI engineer and current Dair.ai builder Elvis Saravia said the focus on performance and cost reduction is exciting, arguing that many other agent management platforms are too expensive and make it difficult to scale agents in production.
Seaotter platform architect and founder Ryan Martin also commented that “Pause-when-idle + governed tools is the ops pattern that scales.” His remark captures the basic operational idea behind the product: keep agents isolated, let them sleep when idle, and tightly control what they can do when they wake up.
How it compares with Docker Cloud Sandboxes
DigitalOcean Managed Agents also arrives at roughly the same time as Docker’s Cloud Sandboxes, and InfoQ notes that the two offerings overlap significantly at the sandbox and runtime level. Both use microVM isolation and persistence, but they are not framed the same way.
Docker Cloud Sandboxes are described as being more focused on developer workflows, especially the ability to move a sandbox between local and cloud compute. DigitalOcean Managed Agents, by contrast, is aimed more squarely at production infrastructure and offers more orchestration and tool support. That difference in emphasis may matter to teams deciding whether they need a portable developer environment or a managed execution platform for agents in production.
What to watch next
Because the service is in public preview, the most important questions now are practical ones: how well the runtime performs under load, how much overhead the tool gateway adds, and how teams will manage permissions as agent fleets grow. The broader AI agent market is moving quickly, but the hard problems remain familiar cloud problems: isolation, state management, security, observability, and cost.
DigitalOcean’s pitch is that it can absorb much of that complexity for customers. If the company can deliver on the promise of managed execution with persistent state and governed tool access, Managed Agents could become a useful option for teams that want to run AI agents as part of real production workflows rather than as isolated demos.
Source: Original report
Was this helpful?
Explore more: DevOps Services More AI & Automation Tech News
Last Modified: October 8, 2026 at 10:33 pm
0 views
