How to turn Claude Co-Work into an AI team

TL;DR

Sub agents in Claude Co-work can run research, planning, and execution tasks in parallel. You can spawn multiple specialized agents, organize outputs in local folders, and persist agent configurations for reuse. The workflow follows four phases: Research, Plan, Execute, QA, with a built in PowerPoint ability and a dedicated content reviewer agent.

Transcript

If you use Claude code, then you probably already know that you can spin up multiple sub agents in parallel to do your work for you. But what most people don't know is that you can do the exact same thing in Claude co-work. So if you're a Claude AI user and you still find the terminal intimidating, then you can get a piece of that firepower just wi... Read More

Key Insights

• Claude Co-work enables parallel sub agents to run research, planning, and QA tasks simultaneously for faster results.
• Agents can be spawned in parallel and managed from the Co-work UI, reducing context window constraints and avoiding sequential bottlenecks.
• A folder structure and markdown-based documentation can persist agent configurations for reuse in future sessions.
• The four phase workflow (Research, Plan, Execute, QA) provides a repeatable pattern to tackle large tasks in Claude Co-work.
• PowerPoint and content reviewer技能 (skills) demonstrate how outputs can be generated and quality checked automatically.
• Users can add or modify agents mid-workflow to adjust scope or address unexpected findings without restarting from scratch.
• The system logs sources and steps for each agent, enabling audit and transparency of the research process.
• Storing agents and outputs in local folders helps maintain continuity across sessions and sharing across teams.

Questions & Answers

Q: What is Claude Co-work and why use sub agents in it?

Claude Co-work is a workspace within Claude that allows you to run multiple specialized agents in parallel. Sub agents are individual agents designed for specific tasks, such as research, planning, or content review. Using them in parallel speeds up complex projects by distributing work across several focused agents, reducing bottlenecks in single-thread execution and preserving context for each task. This setup enables more scalable and modular workflows, making it easier to orchestrate large efforts like building a strategic AI roadmap or preparing multi-part analyses.

Q: How do you start a new sub agent in Claude Co-work?

To start a new sub agent, access the Co-work tab in the Claude desktop app, decide on the tasks you want to parallelize, and select or create the appropriate agent types. You can spawn multiple general purpose or specialized agents, configure them to run in parallel, and assign them to a shared folder for output. You can monitor progress in real time, view each agent’s search steps, and adjust scope if needed without interrupting other agents. This creates a flexible, scalable research and execution pipeline.

Q: What is the four-phase workflow described in the video?

The four-phase workflow consists of Research, Plan, Execute, and QA. In Research, you gather data and sources via parallel agents. In Plan, you organize findings into an actionable roadmap. In Execute, you generate the deliverables such as a PowerPoint deck or documents. In QA, you review outputs for consistency, overlaps, and quality issues, possibly adding a dedicated reviewer agent to automate checks. This pattern keeps tasks structured and auditable.

Q: Can you persist and reuse agents across sessions?

Yes, the video demonstrates persisting agents by storing their configuration and outputs in a local folder, including a README and skill definitions. This allows brand new sessions to quickly load the same agent setups and workflows without rebuilding from scratch. Persisting agents helps maintain continuity for recurring projects, ensures consistency across team members, and saves time by reusing proven configurations.

Q: What roles do the sample agents play in the demo?

In the demo, the sample agents include a general purpose research agent, a plan agent, a content researcher, and a PowerPoint skill. There is also a QA/reviewer role added to check for discrepancies in the final outputs. Each agent has a distinct function, enabling parallel research, structured planning, content extraction, and presentation creation, while a reviewer ensures quality and coherence across slides and text.

Q: How does parallel execution affect the research process?

Parallel execution allows multiple researchers to work simultaneously on different facets of a task, such as emerging AI capabilities, market analysis, and technical validation. This approach speeds up the overall research process and reduces the risk of missing important sources by distributing the workload. It also helps maintain a clear audit trail, as each agent’s steps and sources are visible during execution.

Q: What kind of output can the PowerPoint skill generate?

The PowerPoint skill creates slide decks with design and content generated from the research and planning phases. It can output previews within Claude and may require final adjustments in PowerPoint. The skill is designed to work in tandem with a content researcher and a QA reviewer to minimize text overlaps, font issues, and layout problems, delivering a polished executive presentation for decision makers.

Q: What should you do if you see an unexpected result during execution?

If an unexpected result appears, you can use the Co-work UI to add a question or modify the scope without restarting the entire workflow. This allows for on-the-fly corrections and adjustments to the plan, ensuring that the agents stay aligned with the task goals. By iterating in place, you save time and maintain momentum while preserving the parallel execution benefits.

Summary & Key Takeaways

• The video demonstrates how to activate sub agents in Claude Co-work and spawn multiple parallel researchers, planners, and reviewers, enabling faster complex tasks without overloading the main context window. It emphasizes organizing outputs in a local folder and persisting agent configurations for reuse.

• A four phase workflow is presented: Research, Plan, Execute, QA, with step by step guidance on spawning agents, monitoring parallel tasks, and auditing progress. It also covers error handling and on the fly scope adjustments, ensuring ongoing quality control.

• The demonstration ends with a practical example of creating a slide deck for an AI adoption roadmap, including a PowerPoint skill, a content reviewer, and a readable README to help new sessions reuse the setup.