One of the best reasons to use subagents is that they run in separate context windows. That separation means several can work at the same time without stepping on each other.
Why parallel works
Imagine you are chasing a bug that could live in the API layer, the database layer, or the front end. Instead of investigating one at a time, you can dispatch three investigations at once. Each agent reads its own slice of the codebase in its own window, and each returns a short answer. You get three conclusions back without three piles of file dumps in your main session.
How to ask for it
Describe the split and ask Claude to run the investigations together:
In parallel, have one subagent trace the API request path, one check
the database queries, and one look at the client-side form. Report
what each finds.
Because the built-in Explore agent is read-only, it is a strong fit for parallel investigation. Several Explore agents can search at once with no risk of them changing anything.
Good candidates
- Searching for one pattern across separate modules.
- Gathering facts about unrelated parts of a feature.
- Comparing how three files each solve a similar problem.
Keep it read-heavy
Parallel work shines for investigation, where the agents only read and report. Be more careful running several editing agents at once against the same files, since their changes can collide. Split the work so each agent owns a different area, then bring the results together in your main session and decide what to do.
1 comment
It's a good article