Building an AI Agent Platform with Cloud-Based Virtual Machines and Extended Context

manus · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Cloud-hosted autonomous agent platform

Manus needed to let LLM agents complete long-horizon digital tasks safely and asynchronously, without forcing users to babysit local computer actions or rely on predefined workflows.

Constraints:

Local computer automation can be dangerous and attention-intensive.
Long-horizon agent tasks require dozens of steps before a final answer.
Predefined workflows and multi-role control can limit what a frontier model can improvise.
Browser automation needs enough page context for reliable clicks and decisions.

Architecture

Manus assigns each task a fully functional cloud virtual machine with a file system, terminal, VS Code, and real Chromium browser. A Claude-based agent receives rich context and tools, operates through a long action-observation loop, can use browser text/screenshots/bounding boxes, and runs in the cloud so users can delegate work, leave, and receive results later.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  user[User] --> manus-agent[Manus Agent]
  manus-agent --> cloud-virtual-machine[Cloud Virtual Machine]
  manus-agent --> tool-layer[Tool Layer]
  cloud-virtual-machine --> file-system[File System]
  cloud-virtual-machine --> terminal[Terminal]
  cloud-virtual-machine --> vs-code[VS Code]
  cloud-virtual-machine --> chromium-browser[Chromium Browser]
  chromium-browser --> browser-context-adapter[Browser Context Adapter]
  browser-context-adapter --> manus-agent[Manus Agent]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`User`	Delegates a task and later receives the completed result or notification.	`Manus Agent`
`Manus Agent`	Claude-based long-horizon agent that decides actions from context and observations.	`User`, `Cloud Virtual Machine`, `Tool Layer`, `Browser Context Adapter`
`Cloud Virtual Machine`	Isolated task environment with Linux today and planned Windows/Android environments.	`Manus Agent`, `File System`, `Terminal`, `VS Code`, `Chromium Browser`
`Tool Layer`	Set of VM tools and function calls the agent chooses among while executing a task.	`Manus Agent`, `Cloud Virtual Machine`
`File System`	Persistent workspace where task files can be uploaded, unzipped, inspected, and transformed.	`Cloud Virtual Machine`
`Terminal`	Command-line interface available inside the VM for task execution.	`Cloud Virtual Machine`
`VS Code`	Development environment inside the VM.	`Cloud Virtual Machine`
`Chromium Browser`	Real browser used by the agent for web tasks.	`Cloud Virtual Machine`, `Browser Context Adapter`
`Browser Context Adapter`	Converts browser state into viewport text, screenshots, and bounding-box screenshots for the model.	`Manus Agent`, `Chromium Browser`

Data flow

A user delegates a task to Manus, Manus runs it inside a cloud VM, the agent inspects files, terminal output, browser text, screenshots, and bounding boxes, chooses among VM tools, loops through actions and observations for 30-50 steps when needed, and notifies the user when the task is complete.

“assigned a fully functional virtual machine manus can use the file system terminal vs code and a real chromium browser”transcript, stated

“like in manners an average task will take maybe 30 or 50 steps and then get the final answer. So when we are building manus actually we try every model we can get our hands on but it finds out only sonet can know okay I imaging a very ”transcript, stated

Why this architecture

Manus chose cloud VMs because they isolate risky computer actions and free users from supervising local automation. It chose a minimal framework with rich context because the team believed frontier models handle broad tasks better when given tools and context rather than many predefined workflows or specialized roles.

“running in the cloud it's much safer and also when it's running in the cloud you don't have to pay your attention to it you just you know assign a task to manus you can just close your laptop or just put your phone back to your pocket you ”transcript, stated

“zero predefined workflows it's just a very simple but very robust structure very simple structure But we just left all the intelligence part to the foundational model. At that time it is cloud sol 3.7 and now we have cloud 4 right. So what we are ”transcript, stated

Models used: Claude Sonnet 3.7 · Claude 4

Scale

One Manus user consumed GPU power for two hours per day.
“there's one single user uh he can just consume the GPD GPU power two hours per day just for himself. So I think we can achieve this goal at end of this year and today uh what I want to talk about is about two things uh one is about manners and ”transcript, stated
The initial product was built over five months.
“last October so from last October to this March five months of work. This is what you saw today. Uh which is the manusi in fifth of March. Yeah. So that's the original idea. Yeah. Thank you guy. And also I want to share some details when we building ”transcript, stated
The team tried 42 website use cases without implementing them as predefined workflows.
“more intelligence is that when we release manus actually we put 42 use cases on our website and someone say okay man is is is not a saying you just uh uh predefine some workflows maybe you have like 42 predefined workflows but actually ”transcript, stated
The VM exposed about 27 tools to the agent framework.
“have like 27 tools in the virtual machine. So the agent framework has to decide what's the action should be performed into that virtual machine. So it's it's very important to call the right tool and write the tools parameters uh right. So ”transcript, stated
Manus spent about $1 million on Claude models in the first 14 days.
“this this shirt in a GTC event. Yeah, we spend like $1 million on cloud model in the first 14 days. I think maybe that's why they they invited us here to to give the speed. Yeah, it cost us a lot to be on the stage, you know. Yeah. So, ”transcript, stated

Alternatives considered (roads not taken)

Run automation on the user's local computer.

Rejected because: Local actions can install dependencies or break the user's machine and require continued attention.

“will may be dangerous right after it maybe it will install some dependencies install some software”transcript, stated

Define many predefined workflows for the agent.

Rejected because: Manus wanted a simple structure that lets the foundation model improvise across use cases.

“have like 42 predefined workflows but actually you know inside manus in the core of manus we have zero predefined workflows”transcript, stated

Use a multi-role agent system with specialized agents.

Rejected because: The team viewed role decomposition as control that limits the model's potential.

“You have to specify this a coding agent, this is a search agent, this is blah blah blah agent.”transcript, stated

Run a local Docker browser to reuse local cookies.

Rejected because: The team wanted everything in the cloud so users could stop paying attention to their device.

“actually we we don't have plan for local environment because as I just mentioned we think it's very important to like give the users attention back to users. We don't want to like user to pay attention still to his phone ”transcript, stated

Patterns

`tool-use-orchestration` (agent-platform)

Manus abstracts VM applications into tools and lets the agent choose actions and parameters during long-horizon tasks.

“agent framework has to decide what's the action should be performed into that virtual machine.”transcript, stated

`privacy-preserving-processing` (execution-environment)

Cloud VM execution isolates potentially dangerous software installation or browser actions away from the user's own computer.

“running in the cloud it's much safer and also when it's running in the cloud you don't have to pay”transcript, stated

`context-compaction` (agent-platform)

Browser context is represented as viewport text, screenshot, and bounding-box screenshot so the model can act on web pages.

“The first is the text in this viewport. We will send it to to uh cloud. And the second thing is a screenshot.”transcript, stated

`agent-memory` (agent-platform)

A personal knowledge system stores user preferences such as confirming details before research tasks.

“once you accept that knowledge into your personal knowledge system man will remembers it”transcript, stated

`model-routing` (model-selection)

The team selected Claude models because they handled long-horizon planning, tool use, and computer/browser alignment better than alternatives.

“what we found the best model to run a very long horizon planning sc is cloud solenance models”transcript, stated

Tradeoffs

Execution location

Chose Cloud virtual machines over Local browser or computer automation.

Rationale: Cloud execution protects the user's computer and lets work continue asynchronously.

“assign a task to manus you can just close your laptop or just put your phone back to your pocket”transcript, stated

Agent framework shape

Chose Less structure with richer context over Predefined workflows and specialized role agents.

Rationale: The team believed less control lets frontier models improvise across more use cases.

“then we provide more context to the model and we have less control.”transcript, stated

Browser representation

Chose Viewport text plus screenshots and bounding boxes over A single vision-only or DOM-only browser signal.

Rationale: The agent receives complementary browser context so it can decide what area to click.

“the third thing is a screenshot but with bounding boxes. You know then the model can decide which area he should click.”transcript, stated

Gotchas (what bit them)

Local agents can make destructive changes.

Root cause: Computer-use agents may install dependencies or software on a user's own machine.

Resolution: Run tasks in cloud VMs and notify the user after completion.

“it may break your computer but we think if it's running in the cloud it's much safer”transcript, stated

Many models stop long tasks too early.

Root cause: Chat alignment trained models to answer after one turn rather than continue an action-observation loop.

Resolution: Use Claude-family models that stayed in the long-horizon loop for more steps.

“all other models all of them failed after maybe only one two three iterations”transcript, stated

Tool-call reasoning needed support before native reasoning/tool features matured.

Root cause: The team did not have native reasoning/tool support in earlier Claude versions.

Resolution: Implement a planner-style reasoning injection before function calls.

“before every function call we just will use another specific we call the planner agent”transcript, stated

Evaluation

Manus evaluates its architecture through long-horizon task behavior, model comparison across candidate LLMs, tool-call performance, and production usage/cost observations after launch.

Metrics: Task steps to completion · Model failure after one to three iterations · Tool-call correctness · Daily GPU usage per user · Claude API spend

“for these five months we try all other models all of them failed after maybe only one two three iterations”transcript, stated

“agent product just heavily relied on the tool use on the function calling.”transcript, stated

Outcomes

Manus launched a cloud-VM agent platform after five months of work.
One user consumed about two hours of GPU power per day.
The platform used 27 VM tools and supported long-horizon tasks of roughly 30-50 steps.
The company spent about $1 million on Claude models in its first 14 days.

The transcript cites five months of development, 42 published use cases, 27 VM tools, 30-50 step average tasks, one user consuming two GPU hours daily, and about $1M Claude spend over the first 14 days.

“an average task will take maybe 30 or 50 steps and then get the final answer.”transcript, stated

“we spend like $1 million on cloud model in the first 14 days.”transcript, stated