DigitalOcean launches Managed Agents in public preview, billing per second

A workload that does not fit the VM model

According to DigitalOcean, every era of the cloud has had its own unit of work, and for the last twenty years that unit was the virtual machine hour. Customers paid by the hour whether the machine was working or waiting. The company says agents behave differently: they think in tokens, act in short bursts of compute, and then wait on a model, a tool or a human approval.

An agent might need four copies of itself in the afternoon and none overnight, and it may need its working state back the next morning on another device, or handed to another person or agent. DigitalOcean says the natural units are therefore tokens consumed, microVM-seconds of real execution, and state that outlives the session.

The post names three trends behind the decision. Agentic tasks burn 5 to 30x the tokens of a chat interaction, according to the company. Open-weight models now carry the majority of agent token load, as teams move to them to control unit costs. And teams building agents are struggling with complexity, often combining a harness on a laptop, sandboxes from one startup, inference from two others, storage at a hyperscaler and search from another vendor.

Sandbox, inference and data in one runtime

DigitalOcean describes Managed Agents as three parts working together. The sandbox is an isolated machine to execute code, drive a browser and produce artifacts. Inference serves open-weight and frontier models next to the sandbox. The data layer holds workspace state, checkpoints, databases and retrieval that outlive the session. Handoffs between them are function calls, not WAN hops, the company says.

The DigitalOcean Harness Runtime gives each session its own Firecracker microVM, with its own kernel, hardware-level isolation, sub-second starts and roughly 300ms resumes, according to the post. Sessions continue if a laptop is closed, can be reattached from a phone or handed to a teammate, and a warmed-up session can be checkpointed and forked. Supported agents include Claude Code, Codex CLI, OpenCode, Hermes and LangGraph, or users can package their own agent as a standard OCI image.

The Action Gateway is a managed MCP endpoint to 16,000+ tools across 500+ providers, including GitHub and Stripe as well as a customer's own MCP servers. DigitalOcean says it selects the right tool for an agent's intent, repairs malformed calls, handles rate limits and can require human approval for sensitive actions. Credentials are brokered at execution time and never reach the model, the prompt or the sandbox.

Pricing based on active use

DigitalOcean bills CPU per second of actual use at $0.044 per vCPU-hour and memory at $0.0095 per GB-hour. Snapshots cost $0.05 per GiB-month. The company contrasts this with offerings that bill a VM from boot to shutdown.

As an example, it says a default 2 vCPU, 4 GB session costs $0.126 an hour fully allocated, but about $0.060 at a typical 25% activity level. Auto-pause catches idle sessions, and the CPU line is zero while the agent waits. The company adds that infrastructure cost under a typical run rounds to cents, and that tokens drive most of the spend.

To control token costs, the DigitalOcean Inference Router can send routine steps to hosted open weights such as Kimi K3 and GLM 5.3, and complex reasoning to a frontier model pay-as-you-go, or route through the customer's own key.

What it means for teams building agents

DigitalOcean positions the product against two alternatives: stitching together five or six vendors, which it says creates fifteen integration seams to secure and operate, or using hyperscalers, which it describes as metered across a dozen line items. It says Managed Agents offers a single observability stream covering traces, tool calls, model calls, step timings and cost, with SSO, MFA, role-based access and audit logs as standard features.

The company says a first session needs one CLI command and a spec file, and that users can choose their own harness, framework or model. These are the company's own claims, and the product is still in public preview.