Complexity-Based Task Routing

  • Route tasks to different agents/strategies based on a complexity score
  • Simple tasks (typo fix, quick note) → single fast agent (Haiku/Sonnet)
  • Medium tasks (FR phase, refactor) → standard Claude Code session
  • Complex tasks (multi-phase feature, architecture) → multi-agent pipeline with planning, implementation, and review steps
  • Inspired by Sven’s 7-step workflow where implementation is “autonomous — routing based on complexity score”

Why

  • Not every task needs the same amount of firepower
  • Saves tokens and time on simple tasks
  • Enables better quality control on complex tasks (dedicated review step)
  • Prerequisite for truly autonomous operation

What It Needs

  • Task classifier (estimate complexity from description)
  • Agent definitions per complexity tier
  • Routing logic in src/
  • Model orchestration layer idea was merged into this concept / FR-038

Open Questions

  • How to classify complexity? Heuristic, LLM-based, or both?
  • Should the user confirm the routing, or is it fully autonomous?
  • How does this interact with the existing FR lifecycle?
  • Model orchestration layer (formerly idea-model-orchestration-layer) — merged into FR-038
  • FR-027 (Phone Access) — delivery notifications tie into the “Deliver” step
  • Sven’s 7-step workflow — original inspiration

References