Run an agent in Cowork
Once you're set up, here's how to start a task, point Claude at your team's AI folder, find the right agent, and run it from start to finish.
Thread's agents live as instructions and templates in your team's AI folder. Once Claude can see that folder, running an agent is mostly a conversation: you tell Claude which agent you want, and it walks you through the rest. Here's the full flow.
Step 1 - Start a new Cowork task
Open the Claude desktop app and start a new Cowork task. Give it a clear name if you like, such as "Run recon agent" or the client you're working on. A task is just a workspace for one piece of work, and you can have several going at once.
Step 2 - Point it at your team AI folder
In the task, connect or select your team's AI folder in Google Drive (for example, the Internal AI or AI shared drive). Choose the folder itself, not one file, so Claude can see every agent and all the context inside. Claude reads the folder's instructions automatically, so it picks up how Thread works before you ask for anything.
Step 3 - See which agents you can use
Not sure what is available? Just ask. Type "What agents can I use?" and Claude will list the agents in your folder, what each one does, and when to use it. This is the fastest way to find the right tool for the job, and it's a good habit before you start.
Step 4 - Run an agent
When you know which one you want, tell Claude plainly. For example: "Run the recon agent." Claude opens that agent's instructions and gets ready to work. You don't need to remember the exact steps - the agent already knows its own process.
Step 5 - Give it the information it needs
The agent will ask for whatever it needs to do the job - a client name, a file, a date range, a payroll register, and so on. Answer in the task, or drag in the file it asks for. If you're missing something, tell Claude and it will help you work out the next step. Once it has what it needs, it runs and hands you the result.
Remember: read-first, approval before action
Thread's rule holds for every agent: anything that can change a live system - HubSpot, isolved, email - works read-first and asks for your approval before it creates, changes, or sends anything. Claude shows you what it plans to do and waits for your yes. You are always the one who approves the final action.
Tips for good runs
- Be specific in your requests. The clearer the input, the better the result.
- Start with simpler tasks while you get comfortable.
- If a run doesn't go as expected, stop Claude at any time and try again with more detail.