planner

To plan agent execution, you provide a feature README or architecture document so the planner can select agents, assign models like Opus or Sonnet, and output a structured PLANNER_OUTPUT.md defining parallel phases and dependencies.

Generating an execution plan from an architecture document involves applying project conventions from planner-context.md to select parallel phases, identify dependencies, and establish risk gates for the orchestrator to dispatch work.

Yes, you can use this planning approach to design software architecture from a feature README by determining the appropriate agents, models, and skills needed to build a concrete execution plan.

Yes, you need a separate orchestrator to run agents because the planner itself does not write code or execute agents; it only outputs the PLANNER_OUTPUT.md file that the orchestrator uses to dispatch work.

The planner decides which models to assign for task execution by evaluating the project conventions defined in planner-context.md and selecting the appropriate model tier, such as Opus, Sonnet, or Haiku, for each execution phase.

A limitation of using a meta-dispatcher for workflow planning is that it cannot directly write code or run agents, meaning it strictly requires a downstream orchestrator to act upon the generated PLANNER_OUTPUT.md execution plan.