Build Agent Teams within Claude Cowork in 17 min

Key Concepts

• Claude Co-work: An advanced environment for Claude that allows file access, software stack integration, and workflow automation.
• Agents/Sub-agents: Autonomous entities that perform specific tasks. Sub-agents allow for parallel processing, significantly increasing speed and efficiency.
• Skills: Saved, reusable instructions for specific tasks within Claude.
• Commands: Sequences that string multiple skills together to execute complex, multi-step workflows.
• Connectors (MCP): Tools that link Claude to third-party software (e.g., Apify for web scraping).
• Parallel Processing: The ability to run multiple sub-agents simultaneously to handle bulk tasks.
• Context Window: The "memory" available to an AI; using sub-agents prevents the main agent's context from becoming bloated with raw data.

1. The Power of Agents and Parallel Processing

The core advantage of using agents in Claude Co-work is the ability to move beyond single-task execution. By spinning up sub-agents, users can perform bulk tasks in parallel.

• Efficiency: A task that previously took a long time due to sequential processing can be completed in minutes.
• Context Management: The main agent receives only a summary of the sub-agents' work, preventing the "bloating" of the context window and improving overall performance.
• Scalability: This setup allows for handling hundreds of items (e.g., lead qualification) simultaneously.

2. Real-World Application: Outbound Sales Workflow

The video demonstrates a multi-step sales automation workflow:

1. Lead Qualification: 150 leads are processed by 15 parallel sub-agents to verify ICP (Ideal Customer Profile) criteria (e.g., SEO services, US-based).
2. Lead Enrichment: 18 sub-agents (17 researchers, 1 LinkedIn scraper) gather company size, descriptions, and recent social media posts.
3. Personalized Outreach: 17 sub-agents generate icebreaker messages based on the enriched data.

• Result: 82 leads qualified, enriched, and messaged in a fraction of the time required for manual processing.

3. Framework for Setting Up Workflows

To automate workflows, the following components are used:

• Skills: Define the "how-to" for specific tasks (e.g., "Lead Qualification Skill").
• Agents (Agent.md): Specific instruction files that dictate the behavior of sub-agents.
• Commands: A master instruction (e.g., /outbound-pipeline) that triggers the sequence of skills.
• Connectors: Integration with tools like Apify via MCP (Model Context Protocol) to access data from LinkedIn, Facebook, or other websites.

4. Step-by-Step Implementation

1. Define the Task: Clearly prompt Claude to use parallel sub-agents (e.g., "Spin up 15 parallel sub-agents...").
2. Install Connectors: Use Apify to enable web scraping. Obtain an API key, configure it in Claude’s "Connectors" settings, and enable specific "actors" (scrapers).
3. Create Plugin Structure: Ask Claude to generate a plugin containing:
   • Skills: Individual task instructions.
   • Agent.md files: Specific logic for sub-agents.
   • Command.md: The sequence of operations.
4. Deployment: Export the plugin as a ZIP file or host it on GitHub to import into the "Plugins" section of Claude Co-work.

5. Key Arguments and Limitations

• Sub-agents vs. Multi-agent Teams: The speaker distinguishes between sub-agents (isolated, task-specific, ideal for marketing/sales) and multi-agent teams (communicative, collaborative, ideal for complex engineering tasks like building a C compiler).
• Token Usage: Sub-agents consume a high volume of tokens. Users are advised to use the "Claude Max" plan to avoid hitting limits.
• Scalability Limits: For massive, non-human-in-the-loop workflows (e.g., thousands of leads), the speaker suggests dedicated automation platforms like n8n or Make.com are still superior.
• Optimal Batching: The speaker recommends assigning 5–15 tasks per sub-agent for the best balance of speed and reliability.

6. Notable Quotes

• "Context is one of the most precious resources for these agents, and therefore [sub-agents] will drastically improve the performance."
• "Multi-agent teams are good for tasks where agents have to work together on one shared goal... while sub-agents are preferred when each one can work in isolation."

Synthesis

Claude Co-work has evolved into a powerful automation platform by introducing sub-agents and plugin architectures. By moving from simple prompting to building modular skills and commands, users can automate complex, bulk-processing workflows. While token costs and scale limitations exist, the ability to run parallel, autonomous sub-agents makes Claude a viable alternative to traditional automation tools for many day-to-day business operations.