What is the primary decision rule for determining if a subagent is useful?	Does the intermediate work (the journey) matter to the main thread, or is only the final result needed?
When should subagents be used?	When the exploration can be separated from the execution, and the main thread only needs the final answer.
What is the main benefit of using a subagent for research tasks?	It allows the subagent to perform heavy exploration (reading many files) while keeping the main thread's context clean with a summary.
Why is a subagent better than the main thread for code reviews?	It provides a fresh perspective by running the review in a separate context, free from the history of how the code was written.
When should a custom system prompt be used in a subagent?	When the task requires a specific tone, audience, or style that differs from the model's default technical writing style (e.g., copywriting).
What is an anti-pattern to avoid when using subagents?	Relying on "expert" personas (e.g., "You are a Python expert"), as the model already possesses that knowledge.
When do sequential subagent pipelines fail?	When each step in the pipeline depends on discoveries or outputs from the previous step.
Why is the test runner pattern often discouraged?	Subagents tend to hide necessary information; when tests fail, the full output is needed for diagnosis, not just a "failed" status.
What should be kept in the main thread?	Tasks where the main thread needs to see and react to the steps happening along the way (i.e., when intermediate work matters).
Name two ideal use cases for subagents.	Research/exploration and Code reviews.