Use Cowork Subagents to 10x Your Productivity

One Claude Cowork conversation can only do one thing at a time, and for research-heavy work that gets slow fast. Subagents break that limit: they are workers that Cowork spins up on demand, each running its own task and reporting back while you keep the main conversation moving. In this walkthrough I explain how they work, then run a real market research job with one subagent and then four in parallel.

What you need

• Claude Cowork
• Realistically, a $100 or $200 plan. Using subagents is like having several conversations running at once, which consumes your tokens by a lot.

1. Agents vs subagents (0:00)

First, the distinction. An agent is completely independent: it works by itself. A subagent is different. It needs to be called by Claude Cowork or whatever other tool you are using. In Cowork, triggering one is as simple as saying "Hey, I want you to use subagents."

2. How a subagent actually works (0:29)

When you ask Cowork to use a subagent, it gives that subagent a task. It literally prompts an agent. That agent then browses the web, reads through documents, completes the objective it was given, and returns a response. That response is fed back into your Cowork conversation.

Why does this matter? Take generating a report. Cowork could generate ten reports for you, but it would take a long time because it does them one at a time. A subagent takes the task off Cowork's hands and does it by itself. You can run multiple subagents at once on different tasks, or multiple subagents on the same task. Both patterns are useful, and market research is a perfect example of the first one.

3. Meet the default Cowork subagents (2:01)

Kick things off by asking Cowork what it has available:

Hey, Claude, can you please tell me which subagents you have access to, please?

Cowork ships with about five subagents by default, and you cannot create new ones yet (they are still adding them):

• Explore: finds information when you do not know where it lives
• General Purpose: the workhorse we use in this tutorial
• Plan: helps Cowork plan things out
• Claude Code Guide
• Status Line Setup

The three that matter most are Plan, Explore, and General Purpose. One thing worth knowing: each agent has its own model. Explore runs on Haiku, the smallest Anthropic model, because searching should be quick and does not need deep thought. General Purpose runs on Opus or Sonnet. These agents can also use skills, so each one is its own little version of Claude Cowork.

4. Run your first subagent (3:19)

Now the real work. I asked for a single research task:

Hey, Claude, can you please use one subagent and conduct a very quick market research on the AI landscape in France, please?

You know it is working because Cowork tells you it is running a subagent. One detail I noticed mid-run: I was still on Fable 5 because I run a lot of tests, and the subagent used Fable 5 too. Subagents inherit the model of your main conversation, so if you use Opus, they use Opus.

5. Review the first result (4:35)

Back after the run, I had the full result of the French AI landscape research. You can open all the details of the subagent's work, check every source it used, and even tell it which sources to look for next time.

6. Run four subagents in parallel (4:58)

One subagent is nice. The real gain is parallelism:

Hey, Claude, can you please now use three different subagents, one for Italy, another one for Germany, and a last one for the US? Do the same type of research. Actually, add a fourth one. Let's do Canada as well, and I want you to run them all in parallel.

Cowork started two agents, then fired off the third and the fourth. As soon as two or more subagents are running, you get a very nice visual overview of what each one is doing. Click on any of them and you see exactly which step that subagent is processing right now. This is where the productivity jump happens: you are doing different things in parallel instead of waiting in line.

7. Use subagents for ideation (6:09)

Parallel subagents are not just for research. Say you want to brainstorm the subject line for a cold mailing campaign. Have different subagents each try a different approach to the subject line, then pick the winner from all of their responses. It is a fantastic pattern for ideation.

8. Results, and how far this goes (6:45)

The four subagents came back with a brief summary each. And they can go much further than summaries: they can create documents like Excel files, and they can even build entire web applications for you. You can have different subagents take different approaches to the same build. There is also a more advanced form of this in Claude Code, called specialized agents, which I will cover in an upcoming video.

Pitfalls and tips

• Watch your tokens. Every subagent is effectively a separate conversation. Make sure you are on a $100 or $200 plan before running several at once.
• Subagents inherit your model. If your main conversation runs Fable 5 or Opus, so do the subagents. Check which model is selected before firing off a batch.
• You cannot create custom subagents in Cowork today. You work with the defaults, and General Purpose covers most real tasks.

Where to go next

Subagents get even more interesting when the output is a working app. See how I build live dashboards in Build Live Artifacts in Claude Cowork, or step back and see where Cowork fits in the bigger picture in The Two Agentic AI Tools You Need to Learn.