decompose-with-agents

# Decompose with agents Read [the shared guidance](references/principles.md), then the mode that fits the request: - **Understand existing work:** [mode guidance](references/modes/understand.md). - **Decompose proposed work into criteria:** [mode guidance](references/modes/decompose.md). Use both when needed, keeping findings about the present distinct from proposals for change. Each mode links the relevant templates. Keep a decision/interview log by default; coding work produces specifications and ADRs. Use an existing pair or establish one when delegation is authorized. If separate contexts are unavailable, use the autonomous variant without claiming an independent exchange. Share the actual goal, relevant evidence, constraints, and decision authority, including the parent purpose of a delegated subproblem. One agent probes; the agent holding task context answers, investigates, or makes authorized provisional choices. Discover consequential questions the brief missed. Follow useful branches without manufacturing debate. A valuable question may remain open; neither agreement nor a confident delegate supplies missing user intent. Keep one persistent decision/interview log with a clear writer. Preserve important questions, answers, short reasons, and dependencies. In understanding mode, exchange source-backed findings and distinguish reconstructed purpose. In decomposition mode, refine criteria and a specification; coding work includes ADRs. Consult relevant available skills and repo conventions instead of imposing a competing format. Revisit affected work when an upstream answer changes. End the exchange when the next useful action or experiment is clear, carrying the records into the handoff. Judge its value by consequential gaps exposed or resolved, not question count or apparent agreement.