DigitalOcean puts agent runtimes and 16,000 tools on one managed stack

DigitalOcean opened Managed Agents to public preview, combining Harness Runtime (Firecracker microVM sessions with pause and resume) and Action Gateway (a managed Model Context Protocol endpoint to 16,000+ tools) with its Inference Engine, the service that runs the model, so teams can run coding agents without building the plumbing.

SOFTWARE desk — agent work is bursty; cloud that bills only active CPU and keeps session state across pause/resume is the infrastructure bet under the coding-agent boom.

The product page, also DigitalOcean’s, says the public preview is open to customers who opt in to the preview terms. Its lede and its FAQ say a paused session resumes in about 200 milliseconds. That is not the blog’s about 300 milliseconds, and it is not the 305 milliseconds in the benchmark table. This desk is not averaging them. Each number stays with the page that printed it.

The post’s case for a new runtime is that agent work comes in bursts. An agent checking a checkout error might query logs, run a script, and test a fix, then sit still while it waits on the model, a tool, or a person. The computer still has to keep the files and the working state. A normal virtual machine, the post says, is an empty computer, and teams end up building the pause, the handoff, and the tool locks themselves, while the bill keeps running during the wait. That argument is the company’s. This desk did not time a checkout agent.

Harness Runtime is the execution half. Each session runs in its own Firecracker microVM, with its own processor slice and its own files. A microVM is a very small virtual computer. Firecracker is the software that starts it and walls it off from the next session. Inside that wall an agent can install what it needs and run code. Exec is the call that runs a command in the session. Port forwarding lets a developer preview an app inside the session without putting it on the public internet. Pause takes a snapshot, so files, processes, and context come back on resume. CPU and memory charges stop while the session is paused. Storage that is kept still costs money. Auto-pause fires when the agent is idle, which the post defines as no outgoing calls to a model or a tool. Fork and parallel sessions let a developer split work across subagents. The post says those controls ship as skills inside each supported harness.

The speed numbers are an internal benchmark DigitalOcean dates to 21 Sep 2026, printed as a table on the post. Harness Runtime became ready in 886 milliseconds, under one second, and the first full agent reply took 3.3 seconds. Ready means the microVM is up. The reply includes the model’s turn, so most of that 3.3 seconds is not the virtual machine. Resume-to-ready was 305 milliseconds, about a third of a second. A reply after resume took 2.43 seconds, next to 2.47 seconds for a session that was already running. The post’s point is that waking a paused session did not feel like starting over, in this test. A command inside an already-running session took 189 milliseconds, against 79 milliseconds for Fly.io Sprites. DigitalOcean says its path is 110 milliseconds slower because the command goes through its edge and control plane for sign-in, permission, and an audit trail, and that closing that gap is still a priority. On the same table, Sprites shows create at 483 milliseconds, resume at 145 milliseconds, create-to-reply at 2.59 seconds, resume-to-reply at 2.19 seconds, and an already-running reply at 2.19 seconds. A footnote says Sprites has no resume API. A sprite wakes on the first request, and those two resume figures were derived by subtracting one steady-state command round trip, not read from a resume call. Sessions were asked for at 2 vCPU and 4 GB on every provider. A vCPU is a slice of a processor. 4 GB is the memory in the request. The Sprites guest reported 8 vCPU and 16 GB, so the machines were not the same size. These are DigitalOcean’s measurements. This desk did not rerun them. The model name and the software versions are in Sources.

Action Gateway is the tool half. It is one managed Model Context Protocol endpoint, the standard plug an agent uses to call outside tools, to 16,000+ tools across 500+ providers. The post names GitHub, HubSpot, Stripe, Snowflake, Box, Supabase, and Exa, plus web search, web fetch, browser automation, and DigitalOcean’s own infrastructure APIs. The product page also names PagerDuty. Teams can add their own servers. Credentials are brokered at the moment of the call and, the company says, never reach the model or the sandbox. A sandbox is the walled-off computer the agent is running in. A connection can be an API key, a shared sign-in app, or a per-user sign-in. If a step needs a person to grant access, the gateway hands back a sign-in link and continues after that. Permissions say which tools an agent may use. Sensitive actions can require a human yes. The company says its own tests matched an agent’s intent to a tool 99.3 percent of the time, even when the wording did not match the tool’s name. 99.3 percent is about 993 of 1,000. That figure is internal. It is not an outside audit. The gateway also works with other apps that speak the same plug, not only Harness Runtime.

Supported harnesses on the blog are Claude Code, Codex CLI, OpenCode, Hermes, LangGraph, and a custom agent packaged as an OCI image. OCI means a standard container image, the same kind of package many clouds already run. The product page adds CrewAI to that list. From the command line, the post shows doctl harness-runtime launch as the start. The full command is in Sources. This desk did not run it.

Pricing follows active CPU, billed per second. When the agent is waiting and using no CPU, the CPU charge falls to zero. The post’s example: a session with 2 vCPU averaging 25 percent busy, and a memory peak of 4 GB across an hour, would cost $0.060 for CPU and memory, against $0.126 for a full hour of that allocated capacity. Twenty-five percent busy means the processors were working about a quarter of the hour. $0.060 is a little under half of $0.126. Storage, inference, and tools that are metered separately are extra. Pausing stops CPU and memory charges and keeps the stored state. Action Gateway’s own tools that need a sandbox use Harness Runtime’s compute and memory rates. Third-party tools use their published per-use prices. The product page points to separate pricing pages for Action Gateway and Harness Runtime. This desk did not open an invoice.

Qencode, a media-processing company, is the customer the post describes. Before the agent, the team spent hours each week sorting support requests across Slack, email, and Intercom. The agent reviews each request, judges urgency, sentiment, and client revenue, and creates or updates a Jira ticket, sending low-confidence cases to a person. Early results, the company’s, suggest an estimated 4 to 8 hours a week saved on triage and status reports, and response times moving from several hours to nearly instant. Estimated and early are the post’s words. Murad Mordukhay, CEO and co-founder, is quoted calling it a force-multiplier that gets the right ticket to the right person. OpenHands and Amplitude are named in the lede and not described further. Insights, a view of a run across the runtime, the gateway, and built-in tools, is in private preview. Signals, which the post says will turn runs into feedback for evaluation and reinforcement learning, is coming soon. Reinforcement learning, here, means using a score on this run to shape the next one. Coming soon is not a ship date.

Plain English for the rest of the card: agent = software that takes steps, not only a chat reply. Model Context Protocol = the plug for tools. microVM = a tiny virtual computer. pause = a snapshot that stops CPU and memory billing. resume = waking that snapshot. active CPU = you pay for processor time the agent actually uses. public preview = anyone the company lets in can try it, with the limits of a preview, not a promise that every feature is final.

PRIMARY here: DigitalOcean’s 22 Sep 2026 blog, “Introducing DigitalOcean Managed Agents: One AI-native stack to power your intelligence,” at digitalocean.com/blog/managed-agents-public-preview, and the product page at digitalocean.com/products/managed-agents — Tier A PRIMARY, the company’s own record. The blog page did not print an hour. The public-preview line, the harness examples, the 16,000+ tools, the Inference Engine line, the under-a-couple-of-seconds line, the about-300-millisecond resume line, the OpenHands, Qencode, and Amplitude names, the Firecracker session, the pause and auto-pause rules, the storage-still-billed line, the benchmark table and the 21 Sep 2026 internal date, the Sprites footnote, the machine-size caveat, the Action Gateway providers, the credential rule, the 99.3 percent internal match, the harness list, the pricing example, the Qencode estimate and the Mordukhay quote, Insights in private preview, and Signals coming soon are the blog’s. The opt-in preview terms, the about-200-millisecond resume line, CrewAI, and PagerDuty are the product page’s. NOT claimed: that this desk launched a session, reran the benchmark, matched a tool 99.3 percent of the time, opened an invoice, timed a Qencode week, or counted customers beyond the names the company printed. Distinct from the already-filed starsling-3m-review-runners, rabbit-os3, and subconscious-51m.