Claude Workflows: Beyond Sub-Agents

Summary

Dynamic workflows in Claude Code represent a fundamental shift from simple sub-agent delegation to an autonomous JavaScript orchestration script that manages task decomposition and parallel execution. Unlike traditional sub-agents where Claude holds the plan, workflows empower the script as the orchestrator, enabling massive parallelism with up to 1,000 agents and structured adversarial verification. Key patterns include fan out and synthesize for research, adversarial verification for critical tasks, tournament for creative exploration, generate and filter for volume, and loop until done for iterative refinement. However, the cost can be significant, influenced by agent count times model, orchestration overhead, and verification taxes. Workflows are generally cost-effective for 20 or more independent subtasks, with break-even around five. Crucially, avoid workflows for conversational tasks, dependent subtasks, low-quality needs, or when the output is undefined. A decision framework with four questions—output specificity, subtask independence, correctness vs. speed, and measurable quality gates—helps determine appropriate usage. Understanding this architecture, patterns, economics, and when to abstain is key to avoiding expensive mistakes.