subagent-driven-development

Subagent-Driven Development

Skill Overview

Subagent-Driven Development is a high-quality development workflow that executes an implementation plan by dispatching a fresh subagent for each task and conducting a two-stage review after each task is completed: first a specification compliance review, followed by a code quality review.

Applicable Scenarios

Executing independent tasks with a clear implementation plan

When you already have a detailed implementation plan and the tasks are relatively independent, this workflow ensures that each task is carried out strictly according to the specifications, while the dual review guarantees code quality.

Rapid iteration development within the same session

If you need to execute multiple consecutive tasks in the current session and want to avoid switching sessions or waiting for human intervention, this workflow provides automated review checkpoints to keep development moving continuously.

Projects requiring high-quality code output

When a project has strict requirements for code quality and specification compliance, the two-stage review mechanism—specification compliance plus code quality—ensures that the implementation meets the requirements while maintaining good code quality.

Core Features

Intelligent subagent task assignment

A fresh subagent is dispatched for each task with the complete task context, preventing context contamination. Subagents work in isolated environments, ensuring focus and a higher likelihood of task success.

Two-stage automated review process

The first stage conducts a specification compliance review to confirm that the code fully meets the requirements of the implementation plan. The second stage conducts a code quality review to evaluate maintainability, performance, and best practices. A task is marked complete only after both stages pass.

Flexible model selection strategy

The appropriate model capability is automatically selected based on task complexity: fast models are used for mechanical implementation tasks, standard models for integration and judgment tasks, and the most powerful models for architectural design and review tasks. This ensures quality while controlling costs.

Frequently Asked Questions

When is Subagent-Driven Development suitable?

This workflow is best suited for situations in which you already have a clear implementation plan, the tasks are relatively independent, and you need to execute multiple consecutive tasks within the current session. Compared with manual execution, it provides automated review checkpoints; compared with parallel session execution, it avoids the overhead of switching sessions.

What if a subagent encounters a problem or needs additional information?

A subagent reports one of four statuses: DONE (completed), DONE_WITH_CONCERNS (completed with concerns), NEEDS_CONTEXT (needs context), or BLOCKED (blocked). For NEEDS_CONTEXT, provide the missing information and dispatch it again; for BLOCKED, provide additional context, use a more powerful model, split the task, or escalate to a human, depending on the situation.

What does the two-stage review specifically include?

The first stage is a specification compliance review, confirming that the implementation fully meets the specification requirements, with no missing functionality or unnecessary additions. The second stage is a code quality review, evaluating aspects such as test coverage, code cleanliness, performance, and maintainability. Both stages must pass before proceeding to the next task.