DigitalOcean opens Managed Agents public preview

DigitalOcean opened Managed Agents to public preview on September 22, 2026, after a private preview. Treat it as a cloud harness SKU: Firecracker sessions that pause when idle, plus a governed MCP tool gateway—not a notebook demo or another model launch. That packaging is the layer many teams still assemble themselves from a harness, a sandbox, inference, and tool access.

The investor release names OpenHands, Qencode, and Amplitude as early builders. Product docs last verified September 21, 2026 describe two independently scaling services, Harness Runtime and Action Gateway, now available to all users. Native integration with DigitalOcean's Inference Engine is part of the pitch; the same release says Serverless Inference spans 75+ open and proprietary models.

Idle-pausing microVMs

Each Harness Runtime session runs in a dedicated Firecracker microVM with its own compute and filesystem, plus built-in Chromium and a coding sandbox. Lifecycle APIs persist history and working state, with pause, resume, and fork. Pausing stops CPU and memory charges while retained storage stays billable. Auto-pause kicks in when there are no outgoing LLM or tool calls.

DigitalOcean's own September 21, 2026 benchmark reports the session ready in about 886 milliseconds and resume-to-ready in about 305 milliseconds. Treat those latency numbers as vendor measurements.

Supported adapters include Claude Code, Codex CLI, OpenCode, Hermes, LangGraph, and bring-your-own OCI images.

Governed tools, credentials outside the model

Action Gateway fronts 16,000+ tools across 500+ providers on one managed MCP endpoint. The catalog includes Web Search, Web Fetch, Browser Automation, DigitalOcean infrastructure APIs, and connectors such as GitHub, HubSpot, Stripe, Snowflake, Box, Supabase, and Exa. Teams can add their own MCP servers. Credentials are brokered at execution time and never reach the model or sandbox. Permissions can require human approval for sensitive actions. DigitalOcean's internal testing claims 99.3% intent-to-tool match accuracy.

CPU billing follows actual consumption per second. The investor release lists $0.044 per vCPU-hour and $0.0095 per GB-hour. New DigitalOcean users receive a $5 credit to launch a first session. The offering is a public preview; DigitalOcean says it expects but does not guarantee production-level performance.

Launch a harness from the Cloud Console or doctl, and read the Managed Agents architecture docs before you attach production tools.

Idle-pausing microVMs

Governed tools, credentials outside the model

Sources