Claude Cowork is a desktop application that runs Claude as a persistent AI agent on your computer. Unlike chat, Cowork maintains continuous operation, manages background tasks, and coordinates with other agents and services.

What Is Cowork?

Cowork is the agent runtime for Claude Code. While Code is a CLI for development tasks, Cowork is a desktop application that:

  • Runs agents 24/7 without your interaction
  • Manages multiple parallel tasks
  • Coordinates between different specialized agents
  • Integrates with local files, APIs, and services
  • Provides a dashboard for monitoring agent activity

Think of it as hiring a team of specialists who work continuously on your behalf.

Installation & Setup

Download Claude Cowork

First Launch

  1. Sign in with your Claude account
  2. Select a local directory as your workspace
  3. Configure integrations (Slack, email, etc.)
  4. Define your agents
  5. Create your first workflow

Agent Architecture

Built-in Agents

Cowork includes ready-to-use agents:

Manager — Orchestrates tasks, manages queue, delegates to specialists Researcher — Gathers information, analyzes data, performs due diligence Developer — Codes, tests, deploys, manages infrastructure Writer — Creates content, edits, formats, publishes Analyst — Reviews data, generates reports, identifies patterns

Creating Custom Agents

---
name: customer-support
description: >
  Handles customer support tickets with escalation routing.
  Monitors email, responds to inquiries, flags urgent issues.
model: sonnet
maxTurns: 50
skills:
  - email-response
  - ticket-routing
  - escalation-rules
disallowedTools: [Bash]
---

# Customer Support Agent

Monitors incoming support tickets and responds appropriately.

## Responsibilities

- Read support email inbox every 30 minutes
- Respond to common questions (FAQ-based)
- Categorize tickets by urgency
- Escalate to human team if needed
- Follow up on open tickets

## Escalation Rules

- Account issues → Immediate escalation
- Billing questions → Finance team
- Technical bugs → Developer team
- Feature requests → Product team

Managing Tasks & Queues

Creating Tasks

Tasks are what you want done. You create them, the system assigns them to agents:

# Task: Analyze March Marketing Performance

## Description
Generate a comprehensive analysis of:
1. Website traffic trends
2. Content performance (blog, email, social)
3. Lead quality and conversion rates
4. Competitor activity

## Requirements
- Use data from Google Analytics + Mixpanel
- Compare to February baseline
- Create a visual dashboard
- Executive summary (500 words max)
- Detailed findings document

## Due
March 25, 2026

## Assignee
analyst

Task States

CREATED → ASSIGNED → IN_PROGRESS → REVIEW → COMPLETED
     ↓           ↓              ↓         ↓
  [BACKLOG] [WAITING] [BLOCKED] [REJECTED]
  • CREATED: Task entered the system
  • ASSIGNED: Manager assigned to an agent
  • IN_PROGRESS: Agent is actively working
  • REVIEW: Agent finished, awaiting approval
  • COMPLETED: Approved and delivered
  • BLOCKED: Waiting for external input
  • REJECTED: Didn't meet requirements, sent back to agent

Monitoring Progress

Cowork's dashboard shows:

  • Current task distribution
  • Agent utilization
  • Task completion rates
  • Bottlenecks and delays
  • Quality metrics

Workflows & Automation

Scheduled Workflows

Set agents to run on a schedule:

---
name: daily-standup
schedule: "0 9 * * 1-5"  # 9 AM weekdays
trigger: time
agents:
  - researcher
  - analyst
---

## Daily Standup Generator

Collects updates from all team members and generates a briefing.

## Steps

1. Pull email updates from team
2. Extract calendar events for the day
3. Check project board for changes
4. Summarize blockers and priorities
5. Create markdown briefing
6. Send to Slack #standup channel

Event-Triggered Workflows

Workflows trigger on specific events:

---
name: support-ticket-handler
trigger: email:[email protected]
agents:
  - customer-support
---

## Support Ticket Workflow

Every time a support email arrives, this workflow runs.

Conditional Logic

Route tasks based on conditions:

---
name: invoice-processing
trigger: file:receipts/
agents:
  - analyst
---

## Invoice Processing

1. Analyze invoice for amount
2. If > $5000: Escalate to manager
3. If < $100: Auto-approve
4. Otherwise: Route to finance team

Integrations

Local File System

Cowork watches directories and can trigger workflows:

watch:
  - path: /data/inbox/
    pattern: "*.csv"
    action: import-data

Slack Integration

Send updates, receive commands via Slack:

slack:
  channel: "#ai-operations"
  commands: true
  notifications: true
  reports: daily

Agents post progress updates, and humans can command agents via Slack.

Email Integration

Monitor email addresses and respond automatically:

email:
  accounts:
    - [email protected]
    - [email protected]
  auto-respond: true
  escalation: [email protected]

API Integration

Connect to external services:

integrations:
  - service: stripe
    auth: api-key
    actions: [read-invoices, read-customers]
  - service: github
    auth: oauth
    actions: [read-repos, create-issues]

Agent Orchestration

Manager Agent

The manager agent coordinates all others. It:

  • Receives new tasks
  • Assigns based on agent capabilities
  • Monitors progress
  • Handles escalations
  • Reports completion
---
name: manager
description: >
  Orchestrates all other agents. Assigns tasks,
  monitors progress, handles escalations.
model: opus  # Most capable model for complex decisions
maxTurns: 100
disallowedTools: []  # Full access
---

Agent Communication

Agents communicate via a shared message queue:

{
  "from": "researcher",
  "to": "analyst",
  "task_id": "task_12345",
  "message": "Marketing data collected. See /tmp/march_data.json"
}

Agents can request help from each other:

# In researcher skill:
Analyze this data and summarize trends.
@analyst please review and send to manager with recommendation.

Performance & Monitoring

Agent Metrics

Monitor each agent's performance:

  • Completion rate — % of assigned tasks completed
  • Average duration — Time per task
  • Quality score — Approved vs rejected outputs
  • Utilization — Active vs idle time
  • Error rate — Failed operations

Dashboard

┌─ Cowork Dashboard ────────────────────────┐
│                                            │
│ Active Tasks: 12                           │
│ Agents Online: 5                           │
│ Queue Length: 8                            │
│                                            │
│ Agents Performance                         │
│ ├─ Manager: 100% (5/5 tasks)              │
│ ├─ Researcher: 80% (4/5 tasks)            │
│ ├─ Developer: 60% (3/5 tasks)             │
│ ├─ Writer: 100% (2/2 tasks)               │
│ └─ Analyst: 75% (3/4 tasks)               │
│                                            │
│ Recent Activity                            │
│ ├─ Daily standup completed                │
│ ├─ Marketing analysis in review            │
│ ├─ Feature development blocked              │
│ └─ Support tickets: 12 resolved today      │
│                                            │
└────────────────────────────────────────────┘

Logs & Debugging

View detailed logs for each agent:

# Logs from the manager
cowork logs manager --tail=50

# Logs from specific task
cowork logs task:task_12345

# Search logs
cowork logs --grep "error" --since "24h"

Advanced Features

Dispatch Mode (Research Preview)

Control your desktop agents from your phone. Available on Claude Cowork with a Max subscription (launched March 2026).

Setup takes two minutes:

  1. Open Claude Cowork on desktop
  2. Click "Dispatch"
  3. Scan QR code with Claude mobile app
  4. Type commands on your phone

Dispatch maintains a single persistent conversation between phone and desktop. Your desktop agents execute the task while you're away.

Important: If your desktop sleeps or Cowork closes, Dispatch stops. Success rate around 50% on complex tasks, higher on simple ones (file search, summaries).

Multi-Agent Collaboration

Example: Complex report generation

---
name: quarterly-business-review
agents:
  - researcher   # Gathers data from all sources
  - analyst      # Processes data into insights
  - writer       # Formats findings into report
  - manager      # Reviews and approves
duration: 2 days
---

## Quarterly Business Review

### Phase 1: Research (Day 1)
Researcher collects:
- Financial data (revenue, expenses, metrics)
- Customer feedback (support tickets, surveys)
- Market research (competitors, trends)
- Internal data (team updates, project status)

### Phase 2: Analysis (Day 1)
Analyst processes data:
- Calculates key metrics (growth rate, margin, efficiency)
- Identifies trends and patterns
- Flags anomalies and risks
- Creates visualizations

### Phase 3: Writing (Day 2)
Writer creates:
- Executive summary (1 page)
- Detailed findings (10-20 pages)
- Appendix with data tables
- Presentation slides

### Phase 4: Review (Day 2)
Manager:
- Reviews quality
- Approves for delivery
- Schedules stakeholder briefing

Error Recovery

Agents automatically recover from failures:

retry:
  max_attempts: 3
  backoff: exponential
  on_failure: escalate

escalation:
  support_queue: true
  notification: true

Checklist

  • Install Claude Cowork desktop app
  • Configure workspace directory
  • Set up integrations (email, Slack, APIs)
  • Create custom support agent
  • Define a daily standup workflow
  • Monitor agent performance dashboard
  • Test task escalation and error handling
  • Set up scheduled workflows
  • Configure integrations for your services
  • Review audit logs and agent communication

Sources: