DigitalOcean recently launched DigitalOcean Managed Agents in public preview, offering a managed cloud infrastructure layer for AI agents. The service includes isolated microVM runtimes, governed tool access, and serverless AI inference.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
DigitalOcean has launched Managed Agents in public preview, using two integrated services—Harness Runtime and Action Gateway—to fold isolated agent execution, session persistence, and tool access into a managed layer. The question it aims to answer is not "can agents run," but "can agents run continuously and controllably in production."
DigitalOcean recently launched DigitalOcean Managed Agents in public preview, providing a managed cloud infrastructure layer for AI agents. According to DigitalOcean's statements relayed in the InfoQ report, agentic workflows differ greatly from traditional cloud applications: developers running agents on conventional virtual machines must build and maintain the supporting runtime environment themselves.
This "plumbing work" is spelled out specifically: saving agent context, persisting artifacts and making them accessible beyond the agent that created them, coordinating parallel work, and hardening security for tool access. In addition, standby VM capacity must be maintained so agents can start quickly, and compute resources must be provisioned on demand.
This is precisely the core question this launch aims to answer: when agents move from demo to production, what's missing is not a smarter model, but a managed layer that handles execution, persistence, tool access, and underlying infrastructure on their behalf. It should be noted that the above pain-point description comes from DigitalOcean's own statements, representing the vendor's definition of the problem domain rather than independently verified conclusions.
DigitalOcean Managed Agents consists of two integrated services, Harness Runtime and Action Gateway. According to the report, together they let developers extend what agents can do, while execution, persistence, tool access, and infrastructure are managed by DigitalOcean.
This division of labor is worth calling out separately: Harness Runtime handles "where agents run and how state is kept," while Action Gateway handles "what agents can touch and how they touch it." The product's three core capabilities—isolated microVM runtime, governed tool access, and serverless AI inference—also essentially fall on these two services.
Packaging runtime and tool access into the same layer is the most structurally noteworthy aspect of this launch; as for whether this packaging is less trouble than assembling it yourself under real production loads, the source provides no measured data, so it cannot be judged.
Harness Runtime is built on lightweight microVMs, providing persistent, isolated compute environments for agents. The agent frameworks it supports include coding agents such as Claude Code, Codex CLI, and OpenCode, general-purpose agents such as Hermes, and custom agents built with LangGraph.
At the state level, Harness Runtime can persistently save conversation history and work state across sessions, and sessions can be paused, resumed, or forked. When an agent is idle—that is, when there are no in-progress LLM or tool calls—the session can automatically pause the agent.
At the parallelism level, sessions can be launched in parallel across repositories and tasks, supporting divide-and-conquer, collaborative, and map/reduce workflows. The report also mentions that each session runs on security-hardened compute and storage resources, and allows developers to connect internal services without exposing them publicly.
Looking at "pause when idle" and "sessions can be forked" together, the orientation of this design is to make agent sessions more like schedulable resources rather than processes that must occupy a machine at all times. Seaotter platform architect and founder Ryan Martin, in a comment, called "pausing when idle plus governed tools" a scalable operational model—this is a personal opinion and does not represent overall market feedback.
Action Gateway serves as the interface for agents to access external tools and services through a unified MCP endpoint, currently providing over 16,000 tools, including Web Search, Web Fetch, browser automation, and DigitalOcean infrastructure management APIs, and offering connectors for popular platforms such as GitHub, HubSpot, and Stripe.
Beyond the number, the report emphasizes governance more: the gateway implements centralized permission management to define the tools and operations available to each agent, and requires human approval for sensitive operations. To support multiple agents accessing external systems, the gateway also provides rate limiting, retries, backoff, and timeouts.
This constitutes a clear set of constraints: tool access is constrained by centralized permission management, sensitive operations require human approval; when multiple agents access external systems, rate limiting, retries, backoff, and timeouts are imposed. In other words, the number 16,000 describes the accessible surface, not the range agents can invoke unconditionally.
The source does not enumerate the specific list and availability of these 16,000-plus tools one by one, so this number should be understood as the current scale relayed in the report, not a verifiable complete catalog.
The report notes that DigitalOcean Managed Agents and Docker Cloud Sandboxes appeared close in time and have significant overlap at the sandbox and runtime level: both provide microVM isolation and persistence.
The difference lies in focus. According to the InfoQ author's judgment, Docker Cloud Sandboxes focuses more on developer workflows, especially the ability to migrate sandboxes between local and cloud compute; DigitalOcean Managed Agents focuses more on production infrastructure and provides more orchestration and tool support.
It should be flagged that this comparative conclusion comes from the InfoQ author's judgment, lacks third-party verification, and is an inference rather than a confirmed fact. It is better suited as a reference for understanding the positioning differences between the two, not as a selection conclusion.
Former Meta AI engineer and current Dair.ai developer Elvis Saravia commented on X that the product's focus on performance and cost reduction is exciting, because many other agent management platforms are too expensive and make it hard to scale agents in production. This is an industry insider's personal assessment and constitutes opinion content in the source.
In contrast are the gaps left by the source itself: DigitalOcean Managed Agents is currently only in public preview, with no general availability date announced; pricing information is not disclosed in the source; and the specific technical implementation details of microVM isolation and security hardening are not elaborated.
Therefore, the performance and cost advantages of the public preview currently have no independent benchmarks or measured data to support them. The "cost reduction" Saravia mentions is an expectation and assessment, not a measured conclusion.
There is also a limitation in the information structure: the only source available for this piece is a single InfoQ report, with no first-party product documentation or pricing information provided. This means all product capability descriptions in this article come from this one report's relay, and readers should apply the corresponding discount when evaluating.
Separating the confirmed parts from the unconfirmed parts, the conclusion is actually quite restrained. What is confirmed is the product form: public preview, two integrated services, microVM-isolated runtime, session persistence and pause/resume/fork, unified MCP endpoint, over 16,000 tools, centralized permission management and human approval for sensitive operations, rate limiting, retries, backoff, and timeouts.
What is unconfirmed is whether it is actually cheaper, faster, and better suited for production. The source gives no general availability date or pricing, does not enumerate the tool list, does not elaborate the technical implementation of microVM isolation, and provides no independent benchmarks or measured data.
So the more accurate significance of this launch is: a cloud vendor has turned the most tedious layer in the agent productionization process—runtime, persistence, tool access, and governance—into a managed product, and put it into public preview for testing. Whether it holds up depends on whether pricing, availability, and measured data—these three pieces of the puzzle—can be filled in after the preview period.
All facts in this article come from a single InfoQ report (October 2, 2026, 2-minute read), with no use of first-party product documentation, pricing pages, or independent testing. Product capability descriptions are relayed from the report, among which DigitalOcean's statements about agent development pain points are vendor self-descriptions; the comparative conclusion with Docker Cloud Sandboxes is the InfoQ author's judgment and lacks third-party verification; the comments by Elvis Saravia and Ryan Martin are personal opinions. The public preview stage means general availability timing and pricing have not yet been disclosed, and performance and cost advantages have no independent benchmark support.
DigitalOcean Managed Agents packages the most tedious layer of agent productionization—isolated runtime, session persistence, tool access, and governance—into a managed service, with a clear direction; but at the public preview stage, pricing, general availability timing, and measured data are all still absent. What can be confirmed now is the product form, not its cost-effectiveness.
Primary report
Primary source