In partnership with
In January 2026, Anthropic quietly crossed a major milestone. The company released Claude Cowork, a new AI product designed to autonomously complete real-world tasks.
The intended audience was software engineers. But something unexpected happened in the usage data. Non-engineers started using it.
Boris Cherny, who leads Claude Code at Anthropic, spotted the pattern. Users weren't treating Claude Code as a coding assistant they were treating it as a delegation interface. They gave it a goal and walked away.
The tool designed for programming had become a general-purpose autonomous executor. That behavior was the seed of Claude Cowork take the same agentic engine, strip away the terminal, and make it accessible to anyone who can type a sentence.
The Insight That Changed Claude’s Direction
Claude Code was originally released as a developer tool. Its purpose was straightforward: help engineers write, debug, and understand code faster. Developers started asking Claude Code to organize files, clean folders, plan trips, recover lost photos, and automate everyday tasks. Designers used it to prototype ideas.
Researchers used it to organize data. Inside Anthropic, this behavior stood out. Boris Cherny, who leads Claude Code, noticed that people weren’t just coding with it they were delegating work to it. That insight changed everything.
Built In 10 Days: The Recursive Development Loop
Once the product vision was clear, Anthropic did something that would have been impossible 18 months earlier: they used Claude Code to build Claude Cowork.
The development timeline was split into two distinct phases. In the first two days, the human team did what humans are best at defining constraints, writing the system architecture, setting safety boundaries, and specifying the user experience contract.
From Day 3 to Day 10, Claude Code wrote the application layer: the task planner, the file system adapter, the sandbox permission model, the approval UI, the progress reporting engine, and the integration layer connecting the agent to the host operating system.
By Day 10, a working public preview was live. Today, roughly 90% of Claude Code's own codebase is written by Claude, making it one of the most self-referential software projects ever shipped.
How Claude Cowork Executes Tasks
Cowork is architecturally built on three layers stacked on top of each other. The bottom layer is the Sandbox Runtime an isolated execution environment that constrains exactly which directories and file types the agent is allowed to touch. This is enforced at the OS level, not just by the model's instructions.
The middle layer is the Task Planner the LLM-powered engine that converts a plain-English goal into an ordered, dependency-aware list of atomic steps. Each step has a type (read, write, move, transform, execute), a target (specific file or directory), and a rollback action in case it fails.
The top layer is the Approval Interface the layer users interact with. Before any writes happen, Cowork renders the full plan in plain English for human sign-off. Once approved, the execution engine runs steps sequentially, reporting status after each one.
pseudocode Cowork execution model
# Cowork internal execution model
INPUT: User goal (plain English)
Phase 1 Planning
LLM receives: goal + directory listing + file metadata
LLM outputs: ordered step list with types and targets
Example output:
Step 1: [READ] Scan /Downloads for PDF files matching 'receipt*'
Step 2: [READ] Extract date and total amount from each PDF via OCR
Step 3: [WRITE] Create /Downloads/Tax_2025/ directory
Step 4: [WRITE] Write receipts_summary.xlsx with extracted data
Step 5: [MOVE] Move matched PDFs into Tax_2025/
Phase 2 Human Approval
Cowork renders plan in plain English
User sees: exactly what will be read, written, and moved
User approves or edits before any execution begins
Phase 3 Execution
Each step runs inside the sandbox
Real-time status reported after each step
On failure: agent reads the error, attempts recovery, or pauses for input
Phase 4 Summary
Cowork writes a structured log of every action taken
Rollback instructions included for any reversible operationsThe Technology Stack Behind Claude Cowork
Here's everything known about the technology stack Claude Cowork is built on:
Core Foundation
Cowork is built on the same Agent SDK that powers Claude Code. The underlying LLM is Claude Sonnet/Opus models from Anthropic the same reasoning engine that drives Claude Code.
Sandbox Runtime
Engineers at Anthropic reverse-engineered the app and found Claude uses Apple's VZVirtualMachine framework, downloading and booting a custom Linux filesystem. The tool runs in a sandboxed virtual machine where users define exactly which folders Claude can access.
Connectivity Layer MCP (Model Context Protocol)
The deeper structural layer is MCP the Model Context Protocol which has become the connective tissue allowing Claude to reach into and act upon data across an organization's entire technology stack. MCP-connected Claude can pull context from Slack threads, Google Drive documents, CRM records, and financial systems simultaneously.
Enterprise Connectors
Organizations can connect Claude Cowork to tools like Google Drive, Gmail, DocuSign, and FactSet. Anthropic also shipped new MCP connectors for Google Calendar, Gmail, DocuSign, Apollo, Clay, Outreach, SimilarWeb, MSCI, LegalZoom, FactSet, WordPress, and Harvey.
Runtime Infrastructure
Anthropic acquired Bun an open-source JavaScript runtime to simultaneously accelerate and stabilize Claude Code's performance, which also underpins Cowork.
Example: Turning A Messy Downloads Folder Into A Tax Report
Here's a real-world scenario that captures what Cowork is built for. A freelancer has 11 months of receipts scattered across a Downloads folder: phone screenshots of restaurant bills, PDF invoices from cloud services, scanned paper receipts, and email attachments saved with names like 'IMG_4892.jpg' and 'invoice_final_FINAL_v3.pdf'.
Tax season is approaching. Normally this is a 4-hour manual job. With Cowork, it's a single instruction.
The Instruction
cowork prompt
> Go through my Downloads folder. Find every receipt and invoice from 2025.
> Extract the vendor name, date, and total amount from each one.
> Create a clean Excel spreadsheet with those columns plus a Category column
> (guess the category from the vendor name Software, Food, Travel, Office).
> Move all the source files into a new folder called Tax_Docs_2025.
> Leave everything else in Downloads untouched.The Plan Claude Shows Before Acting
cowork plan display
Cowork Plan ready for your approval:
Step 1 SCAN /Downloads identify files by extension and filename pattern
Target: PDF, JPG, PNG, HEIC files modified in 2025
Expected: ~140 files to inspect
Step 2 READ Each candidate file extract text via OCR (images) or
text parsing (PDFs). Look for: vendor, date, total amount.
Will skip files with no extractable financial data.
Step 3 CREATE /Downloads/Tax_Docs_2025/ directory
Step 4 WRITE /Downloads/Tax_Docs_2025/expense_report_2025.xlsx
Columns: Vendor | Date | Amount | Category | Source File
Estimated rows: 60–90 (after deduplication)
Step 5 MOVE All matched source files → /Downloads/Tax_Docs_2025/
Files NOT matched remain in /Downloads unchanged.
Scope: Read access to /Downloads only
Writes: /Downloads/Tax_Docs_2025/ only
Rollback: Step 5 is reversible. Steps 1–4 create new files only.
Approve this plan? (Yes / Edit / Cancel)The Execution Log (After Approval)
cowork execution log
[Step 1] SCAN /Downloads
Found 312 files total
Candidates (2025, financial): 143 files
Skipped (non-financial): 169 files ✓
[Step 2] EXTRACT from 143 files
PDF text parsed: 61 files
Image OCR: 72 files
No data found: 10 files (skipped, listed in log)
Extracted records: 123 line items
Duplicates removed: 11 (same invoice saved twice)
Final records: 112 ✓
[Step 3] CREATE /Downloads/Tax_Docs_2025/ ✓
[Step 4] WRITE expense_report_2025.xlsx
112 rows written
Categories assigned: Software (34), Food (28),
Travel (19), Office (31) ✓
[Step 5] MOVE 133 source files → Tax_Docs_2025/
179 files remain in /Downloads (untouched) ✓
Task complete.
Time elapsed: 4 min 12 sec
Human time equivalent: ~4 hours
Summary written to: /Downloads/Tax_Docs_2025/cowork_log.txtWhat makes this technically hard without an agent:
This task requires four distinct capabilities in one pass: file classification, OCR on images, PDF text extraction, and spreadsheet generation.
Doing it manually means opening each file, reading it, typing data into a spreadsheet, and moving the file 140 times.
Cowork chains these capabilities in a single plan, handles mixed file formats, deduplicates, and produces a structured output automatically.
Where Claude Cowork Is Most Useful
Expense & Receipt Management : Extract financial data from mixed PDF/image receipts, build spreadsheets, and organize files by tax year automatically
Research Organization : Rename, sort, and summarize hundreds of downloaded papers, notes, and references into a structured folder system
Project Folder Cleanup : Deduplicate files, move assets into correct directories, and rename inconsistently named files across a project
Report Generation : Combine scattered markdown notes, meeting summaries, and data files into a single formatted document
Media Management ; Sort photos and videos by date taken, rename by EXIF metadata, and remove duplicates from camera roll exports
Developer Automation : Batch-rename environment configs, scaffold boilerplate across multiple services, or clean up build artifact directories
Data Pipeline Prep : Normalize CSV exports from different tools into a consistent schema before importing into a database or BI tool
What This Means For Software Development
The 10-day build of Cowork is not just a product story it's a data point about how software development speed is changing. Anthropic's internal numbers are striking: engineers now rely on Claude for about 60% of their work, up from 28% twelve months ago. Internal tools are iterated on dozens of times per day because the cost of making a change has dropped dramatically. The feedback loop between idea and shipped code has compressed from weeks to hours.
The deeper implication is about what human engineers are for in this new model. Claude executes. The most valuable engineering skill in an AI-accelerated team is not writing code fast it's knowing what to build, why, and what guard rails to put around it. Architecture, judgment, and taste become the scarce resources.
Claude Cowork represents a historic shift in technology where AI has begun building itself, compressing a year’s worth of human engineering into a 10-day sprint and transforming the developer from a code-writer into an orchestrator of autonomous agents.
Sponsored Ad
If you enjoy practical AI insights, check out SnackOnAI and support the newsletter by subscribing, sharing, and exploring our sponsored ad—it helps us keep building and delivering value 🚀
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.