Claude Code workflows move toward manager agents and worker agents

Claude Code workflows are moving beyond one long chat with one model into a setup where several agents split a development job. The main pattern uses a Markdown feature spec as the source of truth, then a primary agent called Fable assigns work to several Opus subagents. The primary agent checks their output, sends work back when it needs correction, and can keep a feature build moving for several hours before a final human review.

Related workflows apply the same idea to bug hunting, rule checks, context management, and code review, with several agents looking at the work from different angles. Cost is a major reason this pattern matters. Reported benchmark numbers showed Fable 5 as the manager and Sonnet 5 as the worker reaching 96% of the performance of an all-Fable setup at 46% of the cost; on BrowseComp, that meant 86.8% versus 90.8% accuracy and $18.53 versus $40.56 per problem.

A second pattern, where Sonnet 5 executes and consults a Fable 5 advisor, reached about 92% performance at about 63% of the cost on SWE-bench Pro.