Agentic Coding

Dynamic workflows in Claude Code address three critical failure modes in agentic coding: agentic laziness (incomplete task completion), self-preferential bias (biased self-verification), and goal drift (eroded objectives across long conversations). The solution isolates work across multiple agents with fresh context windows, each handling a single focused task. Claude writes custom orchestration programs on the fly using six composable patterns—classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop-until-done—each structured to prevent one or more failure modes. Key building blocks include agent() for spawning isolated subagents, pipeline() for multi-stage processing, and parallel() for synchronized tasks. Real-world applications span migrations, deep research, verification chains, memory rule adherence (checking CLAUDE.md compliance), root-cause investigation, and exploratory design work. Workflows run in background with visibility through the run browser, support pausing and resuming, and cache completed stages for cost efficiency. The efficiency principle: workflows earn their higher token cost only on tasks genuinely too large, parallel, or prone to self-grading for a single context window; smaller tasks should use single agents to avoid waste.