By Rajesh Beri | May 3, 2026
Most of the enterprise AI news cycle this week belongs to OpenAI's Bedrock landing, Microsoft's Agent 365 GA, and Google's Cloud Next agent platform announcements. Underneath all of that, on April 30, Manus quietly shipped Cloud Computer — an always-on, persistent Ubuntu virtual machine that lives behind every Manus agent, retains files and tools across sessions, and runs continuously whether or not the user is logged in. Three plans (Basic, Standard, Advanced), team plan coming soon, available across web and mobile. No fanfare. No keynote. No GTC stage.
This is the announcement enterprise AI engineering leaders should actually study. Not because Manus is going to take down OpenAI's enterprise share, but because Manus just productized — for non-technical end users — the answer to the single biggest unsolved infrastructure problem in production agent deployments: state persistence. Every serious enterprise agent team has been wrestling with this for nine months. Most have built it themselves on top of E2B, Northflank, or homegrown Firecracker microVM stacks. Manus is now selling the productized version directly to anyone with a browser.
Here is what shipped, why "persistence" is the single word that separates real enterprise agents from demos, how Cloud Computer reframes the competitive map against E2B / Northflank / OpenAI's sandbox / Vercel sandbox, and the playbook for AI engineering and platform teams before this becomes a buy-vs-build conversation in your next quarter.
What Manus Actually Shipped
Cloud Computer is a persistent virtual machine attached to a Manus agent account. It is not an ephemeral sandbox that resets at the end of a chat. It is not a container that disappears when the request finishes. The marketing copy puts it bluntly: "Files Manus creates on it stay there. Tools Manus installs stay installed."
The technical specs, from the launch announcement and Manus's own blog:
- Operating system: Ubuntu Linux. Standard distro, standard package manager, standard tooling.
- Access: SSH from a terminal or a browser-based web terminal in the Manus dashboard. No graphical desktop yet.
- Persistence: 24/7 always-on. Files, databases, installed tools, environment configuration, and credentials persist across sessions. The agent can read and build on files it created hours, days, or weeks earlier.
- Plans: Basic for simple Python scripts. Standard for active websites and APIs. Advanced for team databases. Pricing tied to compute, memory, and storage. Team plan listed as "coming soon."
- Location & storage: User-selectable region and storage size at provisioning time.
- Monitoring: Built-in CPU, memory, and storage dashboards.
- Isolation: Cloud Computer is the cloud-resident environment. A separately announced Manus Desktop option lets the agent interact with local files, with the cloud environment kept isolated as the security boundary.
The use cases Manus is leading with are deliberately mundane — and that is exactly why the product matters. Persistent Slack, Discord, Telegram, and WhatsApp bots. Live databases. Scheduled reporting. Self-hosted open-source tools (WordPress, Metabase, Home Assistant, Odoo). Long-running web scrapers. Internal Python services. None of these work on an ephemeral sandbox. All of them work on Cloud Computer.
The other deliberate choice: the natural-language interface. Manus is selling this to non-technical users who would never provision an EC2 instance, never write a systemd unit, never configure a cron job. Plain English to a Manus agent now provisions a VM, installs the right packages, deploys the bot, and keeps it running. That is a meaningfully different distribution model than E2B or Northflank, which sell to engineering teams.
Why "Persistence" Is the Word That Matters
Inside the AI agent infrastructure conversation, sandboxing has been a solved problem for about eighteen months. E2B, Northflank, Modal, Vercel, Cloudflare, and OpenAI's first-party sandbox all give an agent a contained Linux environment to execute code safely. The unsolved problem — the one that determines whether an agent demo becomes a production system — is what happens between executions.
The taxonomy production teams have been forced to learn:
- Stateless ephemeral sandbox: Spin up, execute, tear down. The default mode for most platforms. Fine for code interpretation, useless for any agent that needs to remember what it did last hour.
- Filesystem persistence: Files survive between executions. Workable for retrieval workflows, breaks down the moment the agent installs a tool or creates a long-running process.
- Full state persistence: Filesystem + memory + running processes + scheduled jobs + open connections. The model Manus Cloud Computer, Northflank's persistent sandboxes, and a small number of homegrown enterprise stacks ship today.
- Snapshot-based: Periodic snapshots of state that can be restored to a fresh environment. Used by some platforms as a compromise between cost and durability.
The session caps tell you which mode each platform actually supports. E2B Enterprise caps sessions at 24 hours and starts at a $3,000/month minimum. Vercel sandbox sessions run 45 minutes to 5 hours. OpenAI's Agents SDK sandbox is execution-scoped. Northflank's persistent sandboxes run until the customer terminates them. Manus Cloud Computer is in the always-on category — explicitly positioned as "no reset, no time limit."
For enterprise AI engineering teams, this is the buy-vs-build axis. Stateless sandboxes are a commodity. Persistent agent infrastructure is the layer where agents actually become production software, and where the cost, security, and operational complexity all live. Manus just made the persistent version available to a Slack channel admin without writing any infrastructure code.
What This Means for the Sandbox Competitive Map
A year ago this market was framed as "sandbox vendors for AI agents," and the conversation was dominated by E2B's $21 million Series A from Insight Partners and the disclosure that 88% of the Fortune 100 had signed up. That framing is starting to fragment. Three buyer profiles are diverging.
The platform engineering buyer — Snowflake, Salesforce, ServiceNow, in-house AI platform teams at large enterprises — wants a per-second priced, securely isolated, programmable sandbox under their own VPC and identity stack. E2B owns this customer. Northflank is the credible challenger for teams that want stateful workloads. The price floor here is enterprise contract minimums in the low five figures per month. Manus is not playing in this segment.
The agent product builder — startups and product teams shipping vertical agents to end customers — needs a sandbox per user session, with strong tenant isolation. This is where E2B's per-second pricing and Vercel's developer-experience are competitive. Manus is not playing here either.
The end-user agent operator — the marketing manager running a Slack bot, the ops lead automating a weekly report, the small business owner self-hosting WordPress, the founder running a personal database of customer notes — has historically been locked out of persistent agent infrastructure by complexity. This is the segment Manus Cloud Computer just opened. The persistent VM is bundled into an existing Manus subscription, the interface is a chat box, and the agent provisions and operates the environment.
The strategic read: Manus is not competing with E2B for Fortune 100 platform deals. Manus is doing for persistent agent VMs what Notion did for databases and what Vercel did for deployment — collapsing infrastructure complexity into a consumer-grade UX, and winning the long tail. That long tail is real. The Slack and Discord automation market, the Metabase / Home Assistant / WordPress self-hosting market, the personal scheduled-job market — these are tens of millions of users who would never adopt E2B but will absolutely pay $20 a month for an agent that "just runs."
The China context matters too. Manus's reported $2 billion acquisition by Meta was blocked earlier this year on agentic AI sovereignty grounds, which is the only reason Manus is still independent and shipping product as Manus rather than as a Meta sub-brand. That regulatory posture means Cloud Computer arrives as an independent platform rather than a Meta feature, and large U.S. enterprises will treat any data-residency or vendor-risk question accordingly. For now, Cloud Computer is best framed as a small-team and prosumer product, not a regulated-industry production runtime.
What AI Engineering Leaders Should Do This Week
Three actions for enterprise AI engineering, platform, and security leaders watching this announcement:
1. Inventory your agent state strategy. Pull a list of every internal agent project — Codex deployments, Salesforce Agentforce builds, Microsoft Copilot Studio agents, custom LangGraph or CrewAI work, Claude integrations. For each, document explicitly: where does the agent's state live between executions? If the answer is "we haven't decided yet," "the agent restarts each session," or "we keep state in the conversation," you have a Cloud Computer-shaped gap that someone in your business unit is about to fill with a SaaS product you do not control. The shadow agent risk that Lenovo's research called out at 82% of enterprises starts here.
2. Run the buy-vs-build math on persistent agent infrastructure. For any agent project that needs filesystem persistence, scheduled jobs, or long-running processes, the choices are: build on E2B's enterprise sandbox plus your own state layer, build on Northflank's persistent sandboxes, build on cloud-native primitives (EKS + EBS + cron), or buy a productized layer like Cloud Computer for the long tail. The math is different for a 50,000-employee bank and a 200-person SaaS company. Most enterprises will end up with both: E2B / Northflank for production-grade regulated workloads, a Cloud Computer-class product for the line-of-business automation tier.
3. Update your shadow AI governance to cover persistent runtimes. Most shadow AI policies today cover model usage and data egress. Very few cover agent persistence. A persistent always-on VM running under an employee's Manus account, with credentials, customer data, and scheduled jobs, is a materially different governance object than a Claude Pro subscription. If your DLP, identity, and offboarding controls do not address "what happens to the always-on agent infrastructure when this employee leaves," now is the week to ask. AI Guard and SPLX-style agent discovery products are starting to cover this surface; the rest of the policy stack lags.
The Bottom Line
Cloud Computer is not the loudest enterprise AI announcement of the week. It will not appear in the Bain or Gartner cycle for another quarter. But persistence is the silent gating function on production agent deployments, and Manus just productized the answer for the long tail of enterprise users — the same long tail that drove shadow IT for fifteen years. Buyer-side AI engineering leaders should treat Cloud Computer as an early signal of what the persistent-agent-runtime category looks like when it is sold as a consumer product. Build, buy, or govern — but do not ignore.
The agents that win in production are not the smartest. They are the ones that remember what they did yesterday.
Sources:
- Manus blog — Introducing Cloud Computer
- Manus launches Cloud Computer with service hosting feature — TestingCatalog
- Cloud Computer Use: agents that run 24/7 in the background — Eyerys
- Manus Cloud Computer: The Always-On Revolution — Barbaros Blog
- How Manus Uses E2B to Provide Agents With Virtual Computers — E2B Blog
- E2B Series A — $21M to Give Fortune 100 Cloud for AI Agents
- Best persistent sandbox platforms for AI agents (2026) — Northflank
- AI Agents in Production: The Sandboxing Problem No One Has Solved — SoftwareSeni
