A Harness for Every Task: Dynamic Workflows in Claude Code

Anthropic blog post announcing dynamic workflows in Claude Code — a feature that lets Claude author custom orchestration harnesses on the fly, written as JavaScript files with special primitives for spawning and coordinating subagents.

The default Claude Code harness runs planning and execution in one context window, which works for coding but breaks down over long-running, parallel, or adversarial tasks. Three failure modes emerge: agentic laziness (declaring partial progress complete), self-preferential bias (favoring its own outputs when verifying), and goal drift (losing fidelity across turns, especially after compaction). Workflows combat these by isolating subagents in their own context windows with focused goals.

The article describes six composable patterns: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop-until-done. It covers use cases from migrations and deep research to triage at scale, evals, and model routing. It also notes when not to use workflows — they burn more tokens, and parallelism must earn its coordination cost. Tips include pairing workflows with /goal and /loop for repeatable runs, setting token budgets, and saving workflows to ~/.claude/workflows or distributing them via skills.

Sources: raw/dynamic-workflows-claude-code Last updated: 2026-08-01