What Cowork Is and How It Differs from Chat

What You Need to Know

Most people discover Cowork after weeks or months of using Claude Chat. They open it expecting a better chatbot. That misunderstanding causes nearly every frustration that follows.

Claude Chat is a conversational interface. You type a question, you get an answer. Upload a file, and Claude reads it and tells you what it found. The work stays with you. Chat is a consultant on a phone call — knowledgeable and articulate, but ultimately just giving advice while you do the actual typing, copying, filing, and formatting yourself.

Cowork is something else entirely. When you open the Cowork tab in Claude Desktop and point it at a folder on your machine, you aren't starting a conversation. You're handing off a job to an autonomous agent. It plans its own approach, executes multi-step workflows, reads and writes files directly on your computer, and — when the job is large enough — spins up parallel sub-agents to finish faster.

How task delegation replaces prompt-response

Chat locks you into a loop: you prompt, Claude responds. Want a different angle? Prompt again. Need a follow-up? Another prompt. You are the project manager, the executor, and the quality checker. Claude is just the advisor.

Cowork flips this. You describe an outcome: "Organise all files in this directory into subfolders by year and month based on their metadata. Rename every image to 'Project-Alpha-[Date].jpg'. If you find duplicates, move them to a folder named 'Review-Required' and do not delete them without my permission." You define what "done" looks like. Claude determines the steps, builds an execution plan, shows it to you for approval, and then does the work.

This mental model shift is the single most important thing the exam tests. Questions will present scenarios where a user micromanages Cowork step by step versus delegating an outcome, and ask which approach produces better results. The answer is always delegation — Cowork's planning engine is designed to optimise the execution path, and prescribing every step actively interferes with that process.

The sandboxed VM: local execution, not cloud processing

When you select a working directory in Cowork, you grant Claude access to that specific folder through a sandboxed VM — a contained environment on your machine. Think of it as a secure office within your building: Claude gets the key to one room but cannot wander the corridors. It reads, edits, creates, and organises files within that boundary. No manual uploading. No downloading results. The changes happen directly in your file system.

Here's the catch that trips up nearly every new user: if you close Claude Desktop or your computer goes to sleep, the task stops immediately. No cloud fallback. The VM needs the app running and the machine awake. A 20-minute file reorganisation job means keeping your laptop open and plugged in — a constraint that doesn't exist in web-based Chat.

Sub-agents and parallel execution

For straightforward tasks — renaming a dozen files, drafting a single document — Cowork works as a single agent. Larger jobs get decomposed. Cowork creates sub-agents: parallel workers that each tackle a portion of the task simultaneously.

Picture processing 100 CSV files. A single agent would grind through them one by one. With sub-agents, Cowork might spawn five parallel workers, each handling 20 files at once — turning a 30-minute job into a 4-5 minute one. The exam expects you to know when parallelisation helps (independent file operations, batch processing) and when it doesn't (sequential dependencies where step 2 needs the output of step 1).

Persistence: Projects and scheduled tasks

Standard Chat conversations are ephemeral. Close the window and the context vanishes — next time, you start from scratch. Cowork changes this with two persistence mechanisms that make it behave more like a colleague with memory than a stateless tool.

Projects group related tasks with dedicated files, context instructions, and memory. A "Q1 Reporting" project remembers your formatting preferences, knows which data sources matter, and carries forward decisions from previous sessions. Domain 5 covers this in depth, but you need to know it exists here because it's a key differentiator from Chat.

Scheduled tasks let you automate recurring work — clean up your Downloads folder every Friday, compile a weekly status report every Monday morning, check your inbox for urgent items daily at 8am. None of this is possible in standard Chat, and it's one of the features the exam tests most heavily in scenario questions.

What the exam really tests

Domain 1 carries 20% of the exam weight — more than any other domain. Don't waste time memorising feature lists. The exam presents realistic business scenarios and asks you to identify the correct approach, spot the architectural misconception, or predict what will go wrong. Roughly 85% of Domain 1 questions sit at application or scenario-analysis level.

The four most tested traps for this task statement:

1. "Cowork processes files in the cloud" — No. Everything runs locally in a sandboxed VM.
2. "Cowork activity appears in Enterprise Audit Logs" — No. Cowork is currently excluded from standard audit logs, the compliance API, and data exports.
3. "Computer Use works on Windows" — No. macOS only in research preview.
4. "Agentic tasks use the same tokens as Chat" — No. Sub-agents, tool calls, and planning overhead mean significantly higher consumption.

Know these cold. They appear in different disguises across multiple questions.

---

Common Mistakes

---

Hands-On Activity

---

Practice Question

---

Sources

• Cowork: Claude Code power for knowledge work — Anthropic
• Get started with Cowork — Anthropic
• Claude Cowork Guide 2026: Skills, Plugins, Connectors & Setup Tips — FindSkill.ai
• Use Cowork safely — Anthropic