154 subagents: feature or smell?

VoltAgent's awesome-claude-code-subagents collects 154-plus specialized Claude Code subagents across ten categories, each defined as a small Markdown file with YAML frontmatter and installable straight into a project. As a reference it's genuinely useful, and the craft is real.

It's also a tell. Specialization is not free: every additional agent is another boundary to design, another activation rule to get right, another thing that can fire when it shouldn't. The quiet evidence is in the categories themselves. One of the ten is "Meta and Orchestration," which is to say a tenth of the catalog exists to manage the other nine. When coordinating the specialists becomes its own specialty, that's the cost showing up on the books.

I'm honestly not sure where the line sits. Context isolation and parallel fan-out justify real specialization: a focused agent with its own tool scope and a clean context window often beats a generalist drowning in instructions. But 154 is well past the point where any one person can hold the system in their head, and a fleet you can't reason about is a liability dressed as capability. My working guess is that most teams want a handful of sharp agents, not a phone book, but that's exactly the question worth pressure-testing. Where have you found the line between useful specialization and a fleet that manages itself more than it manages your work?