Многоагентный pipeline разработки с TDD

orchestration

agentic-workflow is a Claude Code skill that implements a standard five-stage multi-agent pipeline for implementation tasks: research (oracle), planning, validation, implementation, and review. Every agent runs with `run_in_background: true`, and inter-stage handoffs happen via files written to `.claude/cache/agents/<stage>/`, keeping the main conversation context minimal and purely orchestrative. The implementation stage enforces TDD — failing tests are written before any production code, then verified after implementation. The `Task` tool is used instead of `TaskOutput` to avoid flooding the context with 70 000+ token transcripts. Built for developers and teams who need reproducible multi-agent coordination with explicit artifact handoffs and file-system-based progress monitoring.