Task-Specific Custom Harnesses: Dynamic Workflows for Claude Code

Claude Code's new Dynamic Workflows feature allows it to write and execute a harness on the fly, tailored to the task at hand. Unlike existing static workflows, it can run JavaScript files to spawn and orchestrate subagents, and can even decide which model each agent should use and whether worktree isolation should be applied.

The feature is designed to structurally prevent major failure modes that occur within a single context window.

• Agentic laziness: stopping after partial completion instead of finishing a complex task
• Self-preferential bias: a tendency to favor one's own output during verification
• Goal drift: losing sight of the original goal or constraints as turns progress

Claude Code supports a variety of usage patterns, including:

• Classify-and-act: classifying the type of task and then routing it to the appropriate agent
• Fan-out-and-synthesize: breaking a task into smaller steps, running them in parallel, and synthesizing the results
• Adversarial verification: having a separate agent adversarially verify the generated output
• Tournament: having multiple agents compete to select the best solution

This workflow can be used not only for coding tasks but also for a range of non-technical tasks such as research, security analysis, code review, and data analysis.