Subagent-driven development

In standard agentic coding, a single AI session works through an entire codebase change in one long context window. Context accumulates, earlier instructions get buried, and the model starts making decisions based on a noisy history. Subagent-driven development (SDD) breaks that pattern by giving each task its own fresh subagent with precisely crafted context: only what that task needs, nothing inherited from previous tasks.

The workflow starts with a controller agent reading a plan and extracting tasks. For each task, it dispatches a dedicated implementer subagent that writes code, runs tests, commits, and self-reviews. After the implementer signals completion, two review subagents run in sequence: one checks whether the code matches the spec, and a second checks overall code quality. If either reviewer finds issues, the implementer is re-dispatched with specific feedback. The controller never inherits the implementer's messy session history.

SDD also recommends routing tasks to models by complexity: cheap, fast models for mechanical one-to-two file changes, standard models for multi-file integration work, and the most capable models for architecture decisions and final review. This keeps inference costs down without sacrificing quality on the tasks that actually need it. The pattern is popularized by the obra/superpowers framework but applies to any agentic coding setup that supports subagent dispatch.