The course reduces the choice to one question: does the intermediate work matter to the main thread? If only the final result matters, delegate to a subagent; if the main thread must see or react to what is found along the way, keep the work there.1
Where delegation works
| Use case | Why it fits |
|---|---|
| Research and exploration, such as locating JWT validation in an unfamiliar codebase | Many files searched, one location and explanation returned1 |
| Code review | Fresh context without the creation history; the system prompt can encode project review standards1 |
| Tasks that need a different system prompt: copywriting (audience, tone, voice), styling (design-system files loaded first) | The difference comes from the instructions and context, which the main conversation lacks1 |
Anti-patterns
| Anti-pattern | Why it fails |
|---|---|
| Empty expert personas (“Python expert”) | The main conversation already has that knowledge; isolation helps only with a real difference such as a custom prompt, focused context, or controlled tools1 |
| Sequential pipelines of dependent steps (reproduce → debug → fix) | Each handoff compresses away discoveries the next step needs; pipelines suit only independent tasks1 |
| Test-runner subagents | Diagnosis needs full failure output, and a summary like “tests failed” hides it; the course reports this pattern performed worse among the configurations tested1 |
All three are cases of context isolation costing more than it saves.
Related
- Delegation contract: how to specify the task once you decide to delegate.
- Source: Introduction to Claude Code Subagents