A multi-file task coordinates changes across several source files. Treating it like a single-file edit creates a poor tool choice: reserve agent mode for work that must preserve relationships across files, rather than for one trivial change. section 9.2.2
A practical policy starts with one binary decision: use agent mode for multi-file tasks and avoid it for isolated edits. This captures most of the savings without an elaborate usage matrix, while recognizing that autonomous sessions can remain expensive. section 9.2.2
Choose capability according to the task’s coordination burden:
- Use local models for high-volume, simple work such as tab completion and focused single-file tasks. section 9.3.1
- Use cloud frontier models for cross-cutting work, such as refactoring session management across an authentication system or debugging a race condition spanning three services. section 9.3.1
- Expect a capability gap on the hardest tasks: multi-file work requires maintaining a mental model of the codebase and reasoning about architectural trade-offs. section 9.3.1
