Claude's New Workflows: Power, Cost, and When to Use Them

Summary

Ever wondered if your AI is working as efficiently as possible? A new feature called dynamic workflows in Claude Opus allows for massive parallel processing, analyzing all 41 of one user's skills using 5 million input tokens. This feature is distinct from skills, sub-agents, and agent teams, which have different use cases and costs. While skills are reusable recipes, sub-agents are parallel agents that don't share context, and agent teams are small groups that can communicate. Workflows, however, involve Claude writing a JavaScript script to run potentially hundreds of independent agents in parallel, merging their results. This can be incredibly powerful for complex tasks like auditing entire codebases, but it comes with a significant cost, with one user spending half of their two hundred dollar monthly subscription on a single workflow. The new 'ultra code' setting defaults to using workflows, making it the smartest but most expensive option. The key differentiator is often depth versus width: /goal features involve multiple turns until a criterion is met, while workflows execute a broad set of tasks simultaneously. The speaker advises using workflows only when tasks can be broken down into many individual pieces that can run at the same time, and to be explicit in prompts to avoid unnecessary token consumption. A specific example highlighted was the /deep research function, which automatically invokes a workflow for parallel research and generates a cited report.