The Value of Human Expertise in Dynamic Workflows

Regarding Claude Code’s new dynamic workflows. Makes me even more confident that most coding problems benefit from a capable human stepping an agent through a solution to completion.

Dynamic workflows can, “consume substantially more tokens than a typical Claude Code session.” Used without skill, that would be exceptionally expensive and unproductive.

I get the parallelization case in specific conditions. For repetitive tasks with clearly verifiable outputs. Where the user has already established it can perform with minimal review.

Asking many agents to attempt the same problem in different ways and having other agents try to refute the results?

Set based design works that way. Enabling one or more independent design teams to attain an aspirational goal within a known set of practical design constraints. The teams are experts. The constraints are real and well articulated. The aspirational goal is user centered. That’s an investment in enabling human experts to create novel solutions to complex problems.

But it’s a very expensive way to address a solved problem. Approached without the preparation or expertise, it’s testing how many monkeys sitting at typewriters it takes to produce a user manual.