DigitalOcean Managed Agents

DigitalOcean Managed Agents is DigitalOcean's platform for running AI agents in production environments, and it is currently in Public Preview. It consists of two vertically integrated, ready-to-use components: DigitalOcean Harness Runtime and Action Gateway.

1. Purpose The platform is aimed at teams running AI agents in production environments. According to the official description, there are currently three problems with running agents: agents are tied to a single machine (if a local runtime laptop sleeps or the process stops, the session is interrupted, multiple sessions cannot run in parallel, teammates cannot take over, and it is difficult to pause and then resume); infrastructure must be built by the user (after migrating to a cloud VM, session persistence, credential handling, tool integration, and cross-device or cross-team handoff still need to be built and maintained by the user); and tool access lacks governance (each integration requires new credentials, a permission model, and a retry system, and it is impossible to consistently track what the agent did and what the cost was). The goal of Managed Agents is to let users run agents persistently and provide them with the tools needed to complete tasks, without having to build and maintain the surrounding infrastructure themselves.

2. Main capabilities Harness Runtime (Public Preview) is the execution layer. It runs agent sessions in the cloud and keeps them running continuously, with no need for users to configure or maintain infrastructure. Its capabilities include:

• Each session executes in isolation in a dedicated Firecracker microVM, with independent compute and file system;
• Run multiple agent sessions in parallel as managed cloud workloads, including sub-agents and map/reduce-style workflows;
• Persist session state across devices and across team members;
• Pause and resume without losing context, and stop CPU and memory billing during the pause;
• Automatically pause idle sessions (no outbound LLM or tool calls);
• Set checkpoints for sessions and fork new independent sessions from them;
• Bring your own agent as a standard OCI container image and turn it into a reusable template;
• Provide Exec and secure port forwarding APIs for running commands and previewing applications within a session;
• Support scheduled and webhook-triggered automatic runs;
• Provide per-session access control, structured event logs, and observability into token usage and approval activity;
• Support coding-type harnesses (Claude Code, Codex CLI, OpenCode),General-purpose agents (Hermes), as well as agents built with frameworks such as LangGraph.

Action Gateway (Public Preview) is a tool access layer that provides agents with governed access to over 16,000 tools through a single managed MCP endpoint. Its capabilities include:

• Tool discovery and selection matched to agent intent with 99.3% accuracy, matching even when request wording does not match tool names or descriptions, avoiding loading the full catalog into context;
• Credential brokering: tokens are resolved at execution time and never enter the model, agent prompts, or sandbox;
• Support for static API keys, shared OAuth apps, or per-user OAuth; when authorization is required in a workflow, the gateway provides a login link;
• Centralized permission management by tool and operation, with human approval required for sensitive operations;
• Built-in rate limiting management, retries, backoff, and timeouts, scaling with tool traffic;
• Can connect to DigitalOcean-maintained tools (Web Search, Web Fetch, Browser Automation, infrastructure management API), more than 500 third-party vendors (GitHub, HubSpot, Stripe, Snowflake, PagerDuty, Box, Supabase, Exa, etc.), as well as your own MCP servers;
• Can be used with any MCP-compatible application, not limited to Harness Runtime; simply add the endpoint to the application's MCP configuration;
• Provides a Tool Playground for testing configuration, permissions, and policies before production.

III. How to Use Platform and harness are framework-agnostic; users can continue using existing agents without rebuilding around a DigitalOcean-specific harness, and are not locked into a single model vendor. Users can create sessions through the Web console or CLI (doctl) using Claude Code, Codex CLI, OpenCode, Hermes, or LangGraph. Tool access can be configured through DigitalOcean SDK, Inference Engine, or any MCP-compatible client with Action Gateway. At runtime, sessions continue running in the cloud, and after completing the given task, they pause and snapshot their state; afterward, you can reconnect from the same device or a different device and continue from where the agent stopped, with recovery taking about 200 milliseconds. Session creation to first response takes within seconds (officially stated as less than a few seconds).

IV. Platform Integration and Billing Managed Agents can run alongside other parts of the DigitalOcean platform (Inference Engine, Managed Databases, Spaces Object Storage, Kubernetes DOKS, Identity and Access Management), avoiding stitching together multiple vendors, fragmented billing, or separately managing infrastructure. CPU billing is calculated per second of actual consumption, and idle agent costs drop to zero, rather than paying for allocated capacity. Official links to pricing pages for Action Gateway and Harness Runtime are provided.

V. Applicable Scenarios The official list includes two types of teams: first, engineering teams already building or running AI agents (coding agents or general-purpose frameworks) that want to run persistently in the cloud rather than on a laptop or self-managed droplet, and need governed tool access; second, companies building agent-driven products for their own end users, wanting to run tenant-isolated execution, credential brokering, and tool governance underneath the product, without end users needing to see DigitalOcean.

VI. Limitations The platform is currently in Public Preview and requires acceptance of the preview terms before use. Specific features, billing, and availability may change during the preview phase. For pricing details, please refer to the official pricing page.is created to first response within seconds (officially stated as less than a few seconds).

IV. Platform Integration and Billing Managed Agents can run alongside other parts of the DigitalOcean platform (Inference Engine, Managed Databases, Spaces Object Storage, Kubernetes DOKS, Identity and Access Management), avoiding stitching together multiple vendors, fragmented billing, or separately managing infrastructure. CPU billing is calculated per second of actual consumption, and idle agent costs drop to zero, rather than paying for allocated capacity. Official links to pricing pages for Action Gateway and Harness Runtime are provided.

V. Applicable Scenarios The official list includes two types of teams: first, engineering teams already building or running AI agents (coding agents or general-purpose frameworks) that want to run persistently in the cloud rather than on a laptop or self-managed droplet, and need governed tool access; second, companies building agent-driven products for their own end users, wanting to run tenant-isolated execution, credential brokering, and tool governance underneath the product, without end users needing to see DigitalOcean.

VI. Limitations The platform is currently in Public Preview and requires acceptance of the preview terms before use. Specific features, billing, and availability may change during the preview phase. For pricing details, please refer to the official pricing page.