sadd:do-in-steps

Multi-agent orchestration automates complex task decomposition by breaking goals into ordered subtasks, dispatching sub-agents for each step, and verifying progress via an LLM-as-a-judge workflow. Dependencies are analyzed to ensure sequential execution.

The LLM-as-a-judge verification workflow runs a meta-judge and implementation agent in parallel for each subtask. The meta-judge validates progress against consistent criteria before passing context to the next step, ensuring reliable sequential execution.

You need an orchestrator pattern when complex tasks require sequential subtask execution with dependencies. It is essential when models must be selected per subtask and independent meta-judges must verify progress before proceeding to the next step.

To implement parallel meta-judge and implementation agents, instruct the orchestrator to break a goal into steps, run both agents simultaneously per subtask, and validate results using a reusable meta-judge specification with an LLM-as-a-judge.

Yes, sub-agent orchestration is designed for workflows with complex dependencies. It performs automatic dependency analysis during task decomposition and passes context-rich results and decisions between sequential steps to maintain execution order.

A meta-judge workflow requires defining reusable meta-judge specifications for consistent evaluation. While it includes a robust retry mechanism, complex tasks may face limitations if the LLM-as-a-judge verification criteria are not accurately specified for each subtask.