Claude Code Dynamic Workflows: AI Self-Authored Harnesses
Anthropic's Claude Code shipped Dynamic Workflows on May 28, 2026 — enabling Claude to author its own JavaScript execution harness with full structural control: multiple subagents, parallel execution, per-agent model selection, and per-agent isolation. Six named patterns are in active use: fan-out-and-synthesize, classify-and-act, adversarial verification, generate-and-filter, tournament, and loop-until-done. The capability is reaching mainstream practitioner press this week as the "harness is the moat" thesis consolidates across AI newsletters.
Why It Matters
Self-authored execution harnesses mark a structural shift — the agent now controls how it deploys compute, not just what it computes. This is the operational instantiation of the recursive self-improvement loop in day-to-day engineering workflows.