sadd-do-in-steps

Complex multi-step refactoring is automated by decomposing the codebase into sequential subtasks, dispatching specialized AI agents for each phase, and verifying integration with an LLM-as-a-judge. This orchestrates interface updates, implementations, and consumers sequentially while passing context between steps.

LLM-as-a-judge verification uses an independent AI to evaluate the correctness and integration of each subtask completed by specialized agents. It checks agent outputs against requirements before proceeding to the next step, ensuring quality control in automated task orchestration.

Task decomposition breaks down large tasks into manageable, sequential subtasks assigned to specialized AI agents like developers or architects. Each subtask receives relevant context from previous steps, maintaining continuity throughout the workflow automation process.

Iterative refinement is handled through automated retries with feedback up to two times. When an LLM-as-a-judge verification fails, the system automatically retries the failed step with feedback, ensuring robust task completion without manual intervention.

Context passing between sequential AI agents is managed by transmitting relevant information from each completed step to the next. This maintains continuity across the orchestrated workflow, ensuring each specialized agent has the necessary context from prior subtasks.