Hosted AI Agents in 2026: The Cloud Runtimes That Run an Agent for You

Maya runs a two-person candle business off a five-year-old Chromebook that struggles to keep twelve tabs open. Last Tuesday she typed a paragraph into a browser, asked an agent to pull her month of Shopify orders, flag the three suppliers running late, and draft the restock emails, then she closed the lid and picked up her kid. Somewhere in a data center she will never see, a Linux machine spun up, opened a browser, read her store, wrote the emails, and waited. The laptop on her counter did none of that work. It could not have.

This is the quiet revolution in AI that almost nobody is naming out loud. Everyone is arguing about which model is smartest. The thing that changed who gets to use these models is not intelligence. It's location. The agent runs somewhere else now, on hardware you don't own and couldn't afford, and you reach it through a text box.

Call it the hosted agent. It's the unsung hero of the agent era, and it's the reason a person on a budget phone has the same automation muscle as a startup with a rack of GPUs. Here's the full landscape as it stands in 2026: what these things are, who makes them, what they can do, what they cost, and where the honest limits sit.

What a Hosted Agent Is

Start with the definition, because the word "agent" gets stretched to mean everything. A hosted agent is an autonomous AI agent whose model and working environment both run on a provider's cloud, not on your machine. You hand it a goal. A remote computer does the steps. You get the result.

That bundle is the part worth slowing down on. A chatbot rents you the model. A hosted agent rents you the model and a machine for it to work on: a real virtual computer with a browser, a terminal, a file system, and an internet connection. When the agent needs to click a button, it clicks it on that remote machine. When it writes and runs code, the code runs there. Your device is a window, not the workshop.

Contrast this with running an agent locally. A year or two ago, "run your own agent" meant downloading a model, feeding it a GPU, and babysitting a Python script on your own hardware. Powerful, and out of reach for most people. The hosted model flips the arrangement: the intelligence and the machine live in the cloud, and access is the only thing you need to bring. If you've read our take on the shift from generative to functional AI, hosting is the plumbing that made the functional part practical for everyone.

Why This Levels the Playing Field

Here's the claim, stated plainly: hosted agents are the most democratizing thing to happen to automation since the spreadsheet. The reason is boring and enormous at the same time. Compute stopped being something you had to own.

The great equalizer of the agent era isn't a smarter model. It's that the machine doing the work is no longer the machine on your desk.

Think about who this reaches. A freelancer on a secondhand laptop. A teacher with a school-issued Chromebook. A small shop owner whose "IT budget" is a phone. None of them can run a frontier model locally, and none of them need to. A hosted agent gives all of them a Linux box in the cloud that browses, reasons, and works for as long as the task takes, for the price of a subscription or a handful of credits. The person with the cheapest hardware and the founder with the fanciest workstation now reach the same agent through the same text box.

There's a second, sneakier form of leveling. Because the work happens remotely, it keeps going when your device sleeps. Maya closed her laptop and the restock emails still got drafted. That "runs while you're gone" property used to require a server you set up and maintained. Now it's the default. The playing field isn't just leveled on hardware. It's leveled on time, because your agent works a shift you don't have to be awake for. That's the same wonder that makes a repeating task worth handing to an agent in the first place: you describe it once, and it runs without you.

General Agents: Operator, ChatGPT Agent, and Manus

The first camp is the general-purpose web agent: give it a goal in plain language and it drives a browser to get it done. This is where hosted agents went mainstream, and the timeline is worth knowing because it moved fast.

OpenAI lit the fuse with Operator, launched January 23, 2025 and opened to ChatGPT Pro subscribers on February 1. Operator ran a remote browser in OpenAI's cloud and could fill forms, place orders, and book appointments on your behalf. It was a research preview, and its benchmark scores showed the ceiling of that first generation honestly: 38.1% on OSWorld, a test of operating-system tasks, and 58.1% on WebArena, a test of web navigation. Capable, not yet reliable. OpenAI folded it into a bigger product and shut the standalone Operator down on August 31, 2025.

Its successor, ChatGPT Agent, unveiled July 17, 2025, is the clearer picture of where hosted general agents landed. It runs a full virtual machine with a visual browser, a text-based browser, a terminal, and API access, and it can browse, click, type, run code, and download files inside that sandbox. When it hits a login, it hands you a secure takeover view so you type the password yourself and the model never sees it. That combination, a hosted computer plus a human-in-the-loop handoff, became the template everyone copied.

Then came the one that shocked people. Manus, from the Beijing-based startup Butterfly Effect, launched in March 2025 and marketed itself as a general autonomous agent that would take a goal and run with it end to end. It allocates a fully isolated cloud virtual machine for every task, and it went viral on demos of it building sites, screening resumes, and planning trips unattended. Investors noticed: Manus raised a reported $75 million round led by Benchmark in April 2025 at roughly a $500 million valuation, and the company later drew a reported acquisition by Meta at the end of 2025. Whatever happens to the corporate story, Manus proved the appetite: people want to hand over the whole task, not just a sentence.

Hosted Coding Agents: Devin, Jules, and Friends

The second camp is the busiest. Coding is the task hosted agents are best at right now, because software work has clean feedback: the code either runs or it doesn't. These agents live in the cloud, take a ticket, and open a pull request while you do something else.

Cognition's Devin is the poster child, an autonomous "AI software engineer" that works in its own cloud environment with a browser, a terminal, and an editor. Devin bills in compute units rather than seats: an Agent Compute Unit maps to roughly fifteen minutes of active work, so you pay for how long the agent labored, not how many people are on the plan. Seats are unlimited across its plans, which tells you how the pricing logic has shifted. The scarce resource is the agent's time, not the human's.

Google's Jules is the other end of the polish spectrum: an asynchronous agent that opens GitHub pull requests, made generally available on August 6, 2025 and powered by Gemini 2.5 Pro. During its beta, Google reported that developers ran tens of thousands of tasks that produced more than 140,000 public code improvements. Jules gives a free introductory level and then scales limits through Google's AI Pro and Ultra plans, which is a deliberate on-ramp for people who want to try before paying.

The pattern repeats across the majors. GitHub's cloud coding agent lets you assign it an issue, works in an ephemeral GitHub Actions environment, and opens a pull request, and it's multi-model, so you can point it at Anthropic, OpenAI, or Google models. OpenAI ships a cloud Codex agent. Anthropic has pushed Claude Code to run autonomously on managed cloud infrastructure, triggered by a schedule, an API call, or a GitHub event, so it keeps working when your laptop is shut. Replit's agent builds and deploys whole apps from a phone. The common thread: the build happens on their machine, not yours.

One honest note that matters, and it's the same one we make about where no-code hits its ceiling: a hosted coding agent is not always the right tool. If a task is deterministic, a plain script with no model in the loop is cheaper, faster, and more predictable. The good agents, and the good specs, know when to reach for the boring answer.

Enterprise Agent Runtimes

The third camp is quieter and, in revenue terms, the largest. These are the runtimes that host agents for whole companies, wrapped in the identity, governance, and data controls that IT departments require. You don't chat with these. You build on them.

Amazon's Bedrock AgentCore, which reached general availability in late 2025, is the clearest example of the pattern. It's a serverless runtime that hosts an agent built in any framework, LangGraph, CrewAI, or custom code, and wraps it in AWS-grade isolation and networking. The pitch is bring-your-own-agent, and we'll run it safely at scale. Microsoft's Copilot Studio plays the same role for Microsoft 365 shops, and the company has reported hundreds of thousands of organizations building custom agents on it. Salesforce's Agentforce does it for teams whose customer data already lives in Salesforce, with tens of thousands of deals reported since launch. Google, ServiceNow, and IBM all field their own.

Treat those adoption numbers as the vendors' own, because they are, and this series keeps a follow-the-money eye on marketing math. Still, the direction is real. The biggest cloud providers now sell agent hosting as a first-class product, which means the ability to run an autonomous agent is becoming as ordinary as spinning up a database.

How a Hosted Agent Works Under the Hood

You don't need to know the internals to use one, but knowing them tells you why hosted agents behave the way they do. Four ideas cover most of it.

• A disposable virtual machine. Most hosted agents spin up a fresh, isolated computer for each task, do the work there, and tear it down after. That's why Manus talks about a VM "per task" and why coding agents work in "ephemeral" environments. Nothing you do leaks into the next person's session.
• Asynchronous, background execution. The task runs on their clock, not yours. You can close the tab, and the agent keeps going, which is the property that lets it work while you sleep.
• Connectors and tools. The agent reaches your world through pre-built connections to things like Gmail, GitHub, Drive, and Slack, plus the open browser for anything without an integration.
• Human-in-the-loop checkpoints. The good ones pause at logins and high-stakes actions and hand control back to you, so a password stays yours and a "send" or "purchase" waits for a nod.

Put those together and the shape of a hosted agent becomes clear. It's a rented computer with a brain and a leash. The rented computer is why your hardware doesn't matter. The leash is why you stay in control. That's also the exact model behind a good browser agent, which is the most approachable on-ramp to the whole category.

The Timeline and the Numbers

Here's the fast survey in one place: when the notable hosted agents arrived, what they run on, and one fact worth remembering about each. Dates and figures are the providers' own, current as of mid-2026.

Read down the "arrived" column and the story tells itself. This entire category, from first mainstream launch to enterprise-grade runtimes, happened inside eighteen months. That's the pace that makes hosted agents easy to miss and important to understand.

The Trade-Offs Nobody Prints on the Landing Page

Hosting solves the hardware problem. It introduces three new ones, and an honest guide names them.

The bill is metered, so it can surprise you. Credits, compute units, and task quotas all mean the same thing: a long, tool-heavy job costs more than a short one, and a runaway agent can burn through a budget while you're not looking. The fix is to scope tasks tightly and watch the first few runs before you trust a schedule.

Your data takes a trip. The work happens on someone else's servers, so whatever the agent reads, from your inbox to your codebase, passes through their infrastructure. For a candle shop that's fine. For regulated data it's a real question, and it's why the enterprise runtimes exist. Read the data policy before you connect the sensitive system, not after.

Reliability is still uneven. Those Operator benchmark scores, well under human accuracy, weren't a fluke of one product. Hosted agents are genuinely useful and genuinely fallible, especially on long chains and unfamiliar interfaces. The teams getting value from them keep a human on the high-stakes steps and let the agent own the boring ones.

None of this cancels the upside. It just means a hosted agent is a power tool, not a magic wand, and the person who scopes it well beats the person who points it at everything and hopes.

One Spec, Any Runtime

Here's the trap in everything above: each hosted agent is a walled garden. Build your workflow inside Manus and it lives in Manus. Wire it into Bedrock and it's an AWS project now. The runtimes are racing to lock you in while they race to serve you, and both are happening at once.

This is the exact gap BYOBot is built for. You describe the workflow in plain language and get back a portable spec: a structured file that captures the trigger, the steps, the data, and the decisions, independent of any one runtime. Run that spec by pasting it into a chat AI, by handing it to a hosted browser agent, or by wiring it into your automation stack. The spec is yours. The host is a choice you can change.

That portability is the point. Hosted agents leveled the playing field on hardware. A portable spec keeps it level on freedom, so the person on the Chromebook isn't just able to run an agent, they're able to move it when a better or cheaper host shows up. Browse the agent workflow directory to see the shapes these take by task, role, and tool, or read what no-code tools can and can't do for where hosting fits the wider stack. Maya doesn't know or care which data center ran her restock emails last Tuesday. That's the whole idea.

Frequently Asked Questions

• What is a hosted AI agent? +
  A hosted AI agent is an autonomous agent whose model and working environment both run on a provider's cloud servers rather than on your device. You send a task from a browser or phone, and a remote virtual machine does the browsing, coding, and file work, then sends the result back. Operator, ChatGPT Agent, Manus, Devin, and Google Jules are all hosted agents. The defining trait is that the heavy compute lives on someone else's hardware, so you can turn a repeating task into an agent without owning a powerful computer.
• Do I need a powerful computer to run an AI agent? +
  No, and that is the whole point of a hosted agent. The model inference and the virtual computer both run in the provider's data center, so your device only needs to send a prompt and show the result. A five-year-old laptop, a Chromebook, or a phone can command an agent that spins up a Linux machine, browses the web, and runs code for minutes or hours. Running your own agent used to mean owning an expensive GPU. Hosted agents removed that barrier.
• How much do hosted AI agents cost? +
  Most hosted agents bill by usage rather than a flat seat price. Coding agents like Devin meter compute units that map to active working time. General agents like Manus and ChatGPT Agent use credits or task quotas that reset by plan level. Enterprise runtimes bill for the underlying cloud compute the agent consumes. The practical effect is that a long, tool-heavy task costs more than a short one, so the bill tracks how hard the agent worked, not how many people are on the plan.
• Is it safe to give a hosted agent my logins and data? +
  It depends on the provider and how you use it. Because a hosted agent works on remote servers, your data and any pages it visits pass through the provider's infrastructure, so read the data policy before handing over sensitive systems. Most reputable hosted agents add a human-in-the-loop takeover view for logins, where you type the password yourself so the model never sees it, and they pause for confirmation before high-stakes actions. Treat a hosted agent like a capable contractor: useful, but scoped to what you're comfortable sharing.
• Which hosted agent should I use? +
  It depends on the job. For general web tasks, a computer-use agent like ChatGPT Agent or Manus fits. For software work, a coding agent like Devin, Google Jules, or GitHub's cloud agent is the match. For company-wide deployments, an enterprise runtime like Amazon Bedrock AgentCore or Microsoft Copilot Studio aligns with your existing controls. Rather than lock into one, BYOBot writes a portable spec that describes the workflow, so you can run it on whichever hosted runtime fits and change your mind later.