subagent-driven-development

Decompose the task into independent slices with no shared dependencies, then spawn one subagent per slice. After all subagents finish, collect their outputs and reconcile the changes into a consistent result.

A slice is independent when it touches separate files or modules and does not depend on another slice's output. Slices that share state or edit the same files should be merged or sequenced instead of parallelized.

Avoid it for small tasks, tightly coupled changes, or work requiring sequential reasoning where each step depends on the previous one. The coordination overhead outweighs the parallelism benefit in those cases.

Review each subagent's output after collection, identify overlapping or contradictory changes, and resolve them manually or with a follow-up pass. Keeping slices truly independent upfront minimizes reconciliation conflicts.

It requires the ability to spawn subagents plus basic repository access tools: file reading, directory listing, and grep search. These let the orchestrator inspect the codebase, define slice boundaries, and verify subagent outputs.