How to Create Agents in Claude Cowork
Complete guide to building, configuring, and deploying custom AI agents on the Claude desktop platform
Claude Cowork is a powerful desktop environment that enables you to build custom AI agents capable of automating complex workflows directly on your computer. Unlike traditional web-based chat, Cowork gives your agents full autonomy over your local file system, the ability to execute parallel processes, and access to external connectors like Gmail and Google Calendar.
This guide walks through each step of creating, configuring, and deploying agents in Cowork-from initial setup through advanced automation with scheduled tasks and external integrations.
Cowork agents operate in a secure virtual machine with explicit file system permissions. This means your agent can only access folders you explicitly grant access to-providing both safety and flexibility.
1Install the App and Set Up Access
Download the Claude Desktop app from claude.com/download. Cowork operates exclusively in the desktop environment and is not available via standard web chat.
Installation Requirements:
- Install the desktop application
- Sign in using a paid subscription (Claude Pro, Max, Team, or Enterprise)
- Select "Cowork" mode from the interface dropdown
Cowork is available on Mac, Windows, and Linux. If you're working on a Team plan, ensure all team members have desktop app access permissions enabled in your workspace settings.
2Create a Local Workspace
Create a dedicated folder on your computer to serve as your Cowork workspace. Example: ~/Cowork_Projects or C:\Users\YourName\Cowork_Projects
Claude Cowork operates inside a secure virtual machine that works directly with your local file system. This means your agents can only read, edit, and create files in the specific folders you explicitly grant access to. This sandbox model protects your system while giving agents the autonomy they need.
Workspace Structure (Recommended):
3Establish a Project and Core Context Files
For your agent to maintain consistent behavior across multiple sessions, you need to set up a structured Project with context files that define its personality, goals, and operational parameters.
Creating Context Through Interview:
Ask Claude to interview you about your goals, preferences, workflow quirks, and business requirements. This conversational approach generates richer, more aligned context than manual documentation.
Context File Naming Convention:
- agent_goals.md - High-level objectives and success criteria
- agent_personality.md - Tone, communication style, decision-making principles
- interview_context.md - Answers from your interview with Claude
- workflow_rules.md - Domain-specific constraints and best practices
The more explicit your context files, the more consistent your agent's behavior. Include specific examples, edge cases, and failure modes. An agent with weak context will require constant corrections; one with rich context becomes truly autonomous.
4Prompt Claude to Build the Agent Plugin
Your entire agent-its capabilities, workflows, and integrations-can be built entirely through prompt-driven configuration. Simply ask Claude to generate a plugin tailored to your use case.
Sample Prompt:
Core Plugin Components:
A. Skills (Repeatable Workflows)
Ask Claude to save specific repeating workflows as reusable skills. Instead of rebuilding logic each time, your agent references these pre-built functions.
- Skill file naming: skill_email_outreach.md
- Contents: Step-by-step instructions, templates, validation rules
- Reusability: Multiple commands can reference the same skill
B. Agent Instructions (agent.md)
This is your agent's constitution. It defines exactly what the agent needs to do, when to take action, and how to handle edge cases.
- Include specific examples of success
- For bulk workflows: Explicitly tell Claude to configure parallel sub-agents so it processes dozens of tasks simultaneously rather than sequentially
- Define fallback behaviors and error handling
- Specify reporting and logging requirements
C. Commands (command.md)
Commands are the entry points to your agent's workflows. A single command can trigger an entire multi-step sequence by combining multiple skills and agents in order.
- Syntax: Slash commands (e.g., /workout today, /outbound enrichment)
- Composition: String multiple skills and agents together into sequences
- Triggering: Type the command to execute the entire workflow instantly
For high-volume tasks (processing 100+ records), explicitly request parallel processing in your agent instructions. This allows sub-agents to work on multiple tasks concurrently, reducing execution time from hours to minutes.
5Install the Plugin
Once Claude finishes generating your plugin directory, skills, context files, and commands, installation is straightforward.
Installation Options:
- Direct Install: Click the "Save" or "Install" button directly in the Cowork interface
- Manual Upload: Ask Claude to package the plugin as a .zip file, then upload via the "Upload Plugin" button in the plugins menu
After installation, your agent appears as a new plugin in your Cowork sidebar. You can immediately begin interacting with it using the commands you defined.
6Add Connectors and Automate Tasks (Optional)
To unlock your agent's full potential, you can grant it access to external tools and services, then put it on autopilot with scheduled task automation.
Enable Connectors:
Navigate to the "Connectors" tab in your Cowork settings to link your agent to external APIs and services.
- Gmail - Send emails, read inboxes, manage labels
- Google Calendar - Schedule meetings, read availability, send reminders
- Apify - Web scraping and data extraction
- Slack - Post updates, read messages, trigger workflows
- Custom APIs - Integrate with your proprietary tools
Schedule Tasks for Autopilot:
Once connectors are enabled, you can automate recurring workflows without manual intervention.
- Type /schedule or navigate to Scheduled Tasks in your settings
- Paste your automation prompt (e.g., "Check my email inbox every morning at 7 AM and summarize urgent messages")
- Choose a frequency: daily, weekly, hourly, custom cron
- Your agent automatically runs at the specified time without requiring manual triggers
Start with daily schedules to test reliably. Once you confirm your agent handles the task consistently, move to hourly or real-time event triggers for mission-critical workflows.
Key Takeaways
Download Claude Desktop, select Cowork mode, create a local workspace folder.
Interview your needs with Claude, save answers as .md context files that your agent references across sessions.
Prompt Claude to generate agent.md (instructions), command.md (entry points), and reusable skills.
Click "Install" in the UI or upload as .zip. Test with a sample command immediately.
Link agents to Gmail, Google Calendar, Slack, web scrapers, or custom APIs via Connectors.
Use /schedule to run agents on autopilot at specified times (daily, hourly, custom cron).
In a world of noise and routine tasks, Cowork agents represent true delegation. Unlike chatbot interactions where you're always in control, Cowork agents operate with autonomy-checking your email before you wake up, enriching leads while you sleep, analyzing data without you pressing a button. For traders, analysts, and entrepreneurs, this translates to more time for high-impact decisions and less time on mechanical execution. The key is ruthlessness in defining what your agent owns: be crystal clear about its responsibilities, constraints, and success metrics. Vague instructions lead to wasted agent cycles; clear instructions multiply your effectiveness.
This guide was produced by FazDane Analytics based on Claude Cowork documentation. Specific features, menu options, and authentication requirements may evolve with platform updates. Always refer to official Claude documentation for the most current instructions. Cowork agents are subject to rate limits and fair-use policies outlined in your Claude subscription agreement.