DigitalOcean has opened its Managed Agents platform to public preview. The service supplies isolated microVM runtimes for AI agents, governed access to more than 16,000 tools, and tight integration with serverless inference. Developers no longer shoulder the burden of provisioning sandboxes or stitching together authentication layers. They launch sessions and let the platform handle persistence, billing nuances, and security boundaries.
The announcement landed September 22, 2026, after months of private testing. Vinay Kumar, chief product and technology officer at DigitalOcean, framed the shift bluntly. “EC2 was the front door to the first generation cloud,” he said. “Customers rented a virtual machine and assembled everything else around it. The next generation cloud is AI-native, and the agent is the front door.” Business Wire reported the full quote.
That vision sits at the core of Managed Agents. It consists of two tightly coupled services. Harness Runtime spins up a dedicated Firecracker microVM for each agent session. The environment includes its own compute, filesystem, and built-in utilities such as a browser automation layer. Sessions persist conversational history and working state. Teams can pause them when idle, resume in roughly 300 milliseconds, fork new branches, or checkpoint progress. CPU and memory charges halt during pauses. Only active execution draws billing at $0.044 per vCPU-hour and $0.0095 per GB-hour.
Action Gateway operates as the controlled gateway to external systems. It exposes a single managed Model Context Protocol endpoint. Agents discover relevant tools with claimed 99.3 percent accuracy even when phrasing varies from catalog descriptions. Credentials resolve at runtime and never enter the model prompt or the sandbox. Permissions sit under centralized policy. Sensitive actions trigger human approval. Rate limits, retries, and audit logs come built in. The catalog already spans GitHub, Stripe, HubSpot, Snowflake, Supabase, web search, browser control, and DigitalOcean’s own infrastructure APIs. Teams add internal MCP servers at will.
Supported agent harnesses read like a who’s who of current developer favorites. Claude Code, Codex CLI, OpenCode, Hermes, LangGraph, and CrewAI run without modification. Developers also upload standard OCI container images as reusable templates. Sessions launch from the DigitalOcean console or doctl CLI. New users receive a $5 credit to spin up their first session.
Performance numbers released by the company show sessions reaching first response in about 3.3 seconds from cold start in internal tests with Codex CLI and a frontier model. Resume latency hits 305 milliseconds, which the firm says beats leading alternatives by 46 percent. Tool matching accuracy reportedly exceeds conventional methods by 42 percent. Total cost of ownership calculations claim up to 37 percent savings against independent sandbox providers across representative workloads. DigitalOcean’s launch blog details the benchmarks.
Early adopters have begun sharing results. Qencode built a support-triage agent that classifies incoming requests by urgency and sentiment, creates Jira tickets, and escalates only uncertain cases. The team reports weekly time savings between four and eight hours. Amplitude’s CEO Spenser Skates highlighted reduced operational overhead. “This lets our teams spend less time managing infrastructure and more effort on helping customers build better products and get more out of their AI spend,” he stated in the launch materials.
OpenHands also appears among initial builders. The pattern repeats across these cases. Agents now tackle ambitious sequences that cross code execution, data retrieval, external system updates, and collaborative handoffs. A single agent might query logs, reproduce a bug in a sandboxed script, test a patch, open a pull request, and notify a reviewer. Doing so reliably at scale demands durable state, secure tool access, and predictable economics. Previous approaches forced teams to combine self-managed containers, multiple authentication services, separate inference endpoints, and ad-hoc monitoring. Invoices arrived from half a dozen vendors. Costs accumulated during idle periods while agents waited for model responses or human review.
Managed Agents attacks those frictions directly. Isolation comes from hardware-level virtualization rather than container namespaces. No session can reach another customer’s data or compute. Observability surfaces token consumption, approval events, and structured logs in one place. The billing meter stops when agents pause automatically after periods without outgoing calls. Forking lets teams explore multiple solution paths from a common checkpoint without duplicating expense.
Just days after the preview launch, DigitalOcean extended the model. On October 1 it introduced Agent Droplets. These monthly subscriptions bundle runtime, inference credits, tool access, storage, and memory into fixed plans. Pro tier costs $50 per month and applies a 15 percent discount across eligible resources. Team tier runs $200 monthly with 20 percent off. Spending can halt once the allowance exhausts or continue at standard rates. A dedicated Inference and Agents Balance allows prepayment from $5 to $500 that applies only to these workloads. The investor announcement positions the offering as the modern equivalent of the original Droplet, simple enough for an idea at 11 p.m. to reach working software before morning.
InfoQ’s coverage on October 2 underscored the operational headaches the product targets. Maintaining spare capacity for instant starts, configuring compute on demand, and securing credentials across dynamic tool calls have slowed adoption of production agent workflows. By handling the infrastructure layer, DigitalOcean hopes to let engineering teams focus on agent logic and business value. Sergio De Simone’s article notes that sessions support parallel execution across repositories for map-reduce or divide-and-conquer patterns. InfoQ detailed the technical approach.
Documentation highlights additional practical controls. Port forwarding allows previewing applications running inside a session. Webhooks and scheduled triggers automate runs. Per-session access controls and event logs feed into compliance workflows. Security features such as single sign-on, role-based access, audit trails, cloud firewalls, and DDoS protection apply by default.
Yet the service remains in public preview. No service-level agreements govern uptime or performance. DigitalOcean expects the components to reach general availability but offers no guarantees today. Preview terms require explicit opt-in. Production deployments carry risk until formal availability.
Even so, the direction feels deliberate. Cloud providers once sold virtual machines. Then they sold containers and functions. Now the unit of consumption appears to be shifting toward the autonomous agent. DigitalOcean, long focused on simplifying infrastructure for developers and startups, bets that agents will become the primary workload. Its platform integrates inference, execution, memory, and tool governance under one billing and security model.
Competitors offer pieces of the stack. Some provide sandboxed code interpreters. Others specialize in agent frameworks or tool catalogs. Few combine durable microVM sessions, credential-safe gateways, pause-aware billing, and native model access at this level of vertical integration. The 16,000-tool catalog and MCP standardization lower the barrier for agents to act across enterprise systems without custom glue code.
Recent X discussions echo the interest. Developers note that agent runtimes have become infrastructure. Teams experiment with phone-driven oversight while agents continue in the cloud. Cost transparency and governance controls surface repeatedly as must-haves for scaling beyond prototypes. One post highlighted how pausing eliminates charges during model thinking or human approval loops, turning what once looked like unpredictable cloud bills into something closer to measured utility.
Longer term, the success of Managed Agents will hinge on developer experience and economic alignment. If sessions start fast, resume instantly, and expose clear controls for cost and security, adoption could accelerate. Enterprises already running LangGraph or custom coding agents may find the managed layer reduces operational toil enough to justify migration. Smaller teams gain access to production-grade agent infrastructure without hiring dedicated platform engineers.
DigitalOcean has not disclosed exact adoption metrics from the private preview. Customer quotes suggest measurable productivity gains in support triage and internal tooling. The addition of Agent Droplets one week after launch signals confidence that predictable pricing will broaden appeal. Whether the agent truly becomes the new front door to cloud infrastructure remains an open question. But the company has staked territory on that bet with concrete services, benchmarks, and pricing models that address real friction points in current agent deployments.
Teams evaluating the preview can begin through the DigitalOcean console after accepting preview terms. Documentation walks through session creation, tool configuration, and lifecycle management. The platform continues to add connectors and refine performance. For infrastructure teams and AI builders wrestling with fragmented agent stacks, the offering merits close examination. It may not solve every orchestration challenge. It does, however, remove a sizable portion of the undifferentiated heavy lifting that has slowed progress from prototype to production agent workflows.
DigitalOcean Positions Agents as the New Cloud Front Door With Managed Agents Preview first appeared on Web and IT News.