The architecture map of Claude Cowork
Cowork looks like a chat, but it's built like an agent platform: an isolated Linux machine, a planner that breaks down tasks, and cloud sessions that keep running while your laptop is closed. The full map, plus three workflows that make use of it.
What you get
A map of Cowork's 4 architecture layers, what each layer actually means in practice, and three workflows that push it all the way
Who it is for
For anyone who opened Cowork, sent a message, got an answer, and works with it like just another chat. Underneath the interface is a full architecture that changes how you should work.
1. Why it's worth lifting the hood
Most people opened Claude Cowork, sent a message, got an answer, and sent another one. That's how you work with a chat. Cowork is not a chat.
- Underneath the simple interface is a full architecture: an isolated Linux machine, a planner that breaks down tasks, and sessions that run in the cloud even when your computer is off.
- Someone who understands the architecture uses the tool differently. That's the point of this map: show what's actually happening under the hood, then give you 3 workflows that push it all the way.
- Everything here is based on Anthropic's official documentation, as of July 23, 2026. Sources at the bottom.
2. The diagram
┌────────────────────────────────────────────────────┐
│ Claude Cowork │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ (1) Agent Loop │ │
│ │ Chat, file reads, writes, web access │ │
│ │ Runs natively on your device │ │
│ │ Permission gate on every action │ │
│ └───────────────┬──────────────────────────────┘ │
│ │ │
│ Only code and shell commands go down here: │
│ ▼ │
│ ┌──────────────────────────────────────────────┐ │
│ │ (2) Isolated virtual Linux machine │ │
│ │ One per session │ │
│ │ macOS: Apple Virtualization │ │
│ │ Windows: Hyper-V │ │
│ │ Outbound network filtering + OS isolation │ │
│ └──────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ (3) The Planner │ │
│ │ Plan -> break into subtasks -> │ │
│ │ run independent parts in parallel │ │
│ └──────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ (4) Cloud session (beta) │ │
│ │ Its own sandbox at Anthropic, per session │ │
│ │ Keeps running when your laptop is closed │ │
│ │ Scheduled tasks with no device turned on │ │
│ └──────────────────────────────────────────────┘ │
│ │
│ Underneath it all: the same agent engine │
│ that runs Claude Code, without a terminal │
└────────────────────────────────────────────────────┘3. Layer 1: The Agent Loop
This is the part you see: the conversation, reading files, writing files, network access. This layer runs natively on your device, not inside the virtual machine.
- What protects you here is the permission gate: Cowork only touches folders you explicitly connected to it, and every sensitive action goes through an app-level permission system.
- What this means in practice: only connect Cowork to the folders that are relevant to the task. That is good for security, and it also keeps the agent focused.
4. Layer 2: The isolated Linux machine
Every piece of code and every shell command that Cowork runs in a local session does not run on your computer. It runs inside a dedicated virtual Linux machine, isolated from your operating system by the platform's hypervisor: Apple Virtualization on Mac, Hyper-V on Windows.
- The machine is created at the session level. One session equals one machine, with outbound network traffic filtering, OS-level restrictions, and per-session user isolation.
- What this means in practice: the fear that "it'll run on my computer and break something" is based on an old model. The code runs in a closed box. Your files are only accessible through layer 1, and only in the folders you approved.
- To be precise: the isolation is per session, not per subtask.
- In an enterprise environment, note: Anthropic states that Cowork activity is not currently recorded in enterprise audit logs, and EDR tools cannot see what happens inside the machine. That's an isolation advantage, not a monitoring advantage.
5. Layer 3: The Planner
Before Cowork touches a single file, it builds a structured plan and breaks the work into subtasks. When parts don't depend on each other, it runs them in parallel.
- In Anthropic's own words: large projects split into parts that run together.
- What this means in practice: when you feed Cowork small messages, you're actually doing the planner's job yourself, more slowly. Give it the whole project with full context, and let the breakdown happen on its end.
6. Layer 4: The cloud session
A remote session does not run on your device at all. It gets its own sandbox on Anthropic's infrastructure, created when the session starts and destroyed when it ends, with no shared state between sessions.
- The result: close your laptop, the work keeps going. Schedule a task, and it runs even when none of your devices are on.
- What this means in practice: this is what turns Cowork from a tool that waits for you into a tool that works for you. The three workflows below are built on this layer.
- To be precise: this capability belongs to cloud sessions, currently in beta. A classic local session still depends on your device.
- And underneath it all: according to Anthropic, Cowork uses the same agent architecture that runs Claude Code, without a terminal. If you already work with Claude Code, your instincts about an agent that plans, runs, and verifies apply here too.
7. Workflow 1: Fire and close
Uses layer 4 (cloud session) and layer 3 (the planner). At the end of your workday, open a cloud session and give it one complete task with full context. Close the laptop and go.
- Phrase the task as the final deliverable, not the first step. The agent plans the steps on its own.
- Attach all the context up front: files, links, and a definition of "what counts as done."
- Open it in the morning, review the output, fix it with one focused message.
- Why it works: the planner breaks the work down on its own, the cloud doesn't need you awake, and you only step in at the decision point.
Go through every pricing file in the folder, build an organized comparison, and prepare a draft document with a recommendation.8. Workflow 2: Long, scheduled tasks
Uses layer 4 (scheduled tasks) and layer 2 (isolation). Good for tasks that repeat on a fixed cadence: a weekly summary of a reports folder, cleaning up and organizing files, tracking changes and updating a central document.
- Set a scheduled task at a cadence that suits you (daily, weekly).
- Fix a consistent output format, so you can compare each run to the last one.
- Once a week, review the output and update the instructions, the way you'd update an employee.
- Why it works: the task runs without a device turned on, in an isolated environment, and you get a consistent output without having to remember to run it.
9. Workflow 3: Parallel tracks
Uses layer 3 (breakdown and parallel execution). Have a project with several independent parts? Don't open a separate conversation for each part, and don't feed it piece by piece. Give it the whole project and state the breakdown.
- State explicitly which parts don't depend on each other.
- Ask for a separate, clear deliverable for each part.
- At the end, ask for a merge: "Now combine the three deliverables into one document."
- Why it works: you hand the planner the map, and it runs whatever it can together instead of one after another.
Three independent parts: competitor research, a landing page draft, a pricing table. They can be worked on in parallel.10. Where it's available now
As of July 23, 2026. This is an active rollout, so things can change fast. Check the current status at the architecture page in the official documentation (in the sources section).
- Desktop (Mac + Windows): fully available since April 9, 2026.
- Subscriptions: included in every paid plan, Pro, Max, Team, and Enterprise. Not available on the free plan.
- Web and mobile: still in beta, rolling out gradually, first to Max subscribers.
11. Sources
Every fact in this map was checked against Anthropic's official sources: the architecture overview, the getting-started guide, the product page, the release notes, and the pricing page.
https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview
https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork
https://claude.com/product/cowork
https://support.claude.com/en/articles/12138966-release-notes
https://claude.com/pricingHow to use this now
The change is not another feature, it's a work habit: instead of feeding Cowork small messages, give it a whole project with full context, and let the planner, the sandbox, and the cloud do what they're built to do.