DigitalOcean launched Managed Agents in public preview, with a Firecracker microVM runtime for coding agents like Claude Code and Codex, and a single MCP endpoint that reaches 16,000 tools

A team that runs a coding agent on a cloud virtual machine has been writing the same scaffolding every time: session storage, credential handling, parallel runs, an idle-pause so the bill does not run while the agent waits for a model reply. DigitalOcean put a managed version of all of that into public preview on 1 October 2026.

Managed Agents is two services. Harness Runtime runs each agent session inside a dedicated Firecracker microVM with its own compute and filesystem. Action Gateway gives the agent one managed MCP endpoint onto 16,000 external tools from more than 500 providers. OpenHands, Qencode and Amplitude are listed as early users.

Which agents run on it

DigitalOcean names the coding agents Claude Code, Codex CLI and OpenCode, the general-purpose agent Hermes, and agents built with LangGraph and CrewAI. A team can also package its own agent as an OCI container image and use it as a reusable template, so there is no DigitalOcean-specific harness to rebuild around.

Each session persists conversation history and working state across devices and teammates. A session can be paused, resumed, forked, or run alongside others. Auto-pause starts when the agent stops making calls to a model or a tool, and CPU and memory charges stop at that point. Stored state keeps billing.

The tool gateway

Action Gateway covers DigitalOcean's own tools, including Web Search, Web Fetch, Browser Automation and the DigitalOcean infrastructure APIs, plus connectors to GitHub, HubSpot, Stripe, Snowflake, PagerDuty, Box, Supabase and Exa, and a team's own MCP servers.

DigitalOcean says credentials are brokered at execution time and never reach the model, the prompt, or the sandbox. Static API keys, shared OAuth apps, and per-user OAuth are all supported, with mid-workflow sign-in links when authorisation is needed. Rate limits, retries, backoff and timeouts are built in. Human approval can be required on sensitive actions.

DigitalOcean reports an internal test in which the gateway matched the agent's intent to a tool with 99.3 percent accuracy, even when the request's wording did not match the tool name or description. That number is DigitalOcean's own and the comparison is to conventional lexical search within the same product.

Pricing and benchmarks DigitalOcean reports

A session with two vCPUs averaging 25 percent CPU use and a 4 GB memory peak over an hour costs $0.060 in CPU and memory, against $0.126 for a full hour of that allocated capacity. Storage, inference and separately metered tools are extra.

DigitalOcean reports its own startup benchmark on Harness Runtime: the runtime is ready in 886 milliseconds and the first agent response arrives in 3.3 seconds. Resume to readiness takes 305 milliseconds, and a resumed session answers in 2.43 seconds, close to the 2.47 seconds for a session that was already running. These are DigitalOcean's own measurements against its public API, running the same coding agent and model.

Command execution is the one place DigitalOcean reports it is behind: 189 milliseconds per command in Harness Runtime, against 79 milliseconds in Sprites, a difference the post attributes to routing through its edge and control plane for authentication and audit.

The plans are a free tier with a $5 credit, Pro at $50 a month, and Team at $200 a month.

Source

DigitalOcean, Introducing DigitalOcean Managed Agents, 1 October 2026. Also InfoQ, DigitalOcean Managed Agents Brings Managed Cloud Infrastructure to AI Agents, by Sergio De Simone, 2 October 2026.