🤖 I Built 100 Claude Code Subagents. These Are The 12 That Actually Earn Their Context.

Short summary

After testing 100 Claude Code subagents across real engineering work, the author distilled them to 12 keepers. The critical misconception: subagents aren't AI personalities but context firewalls that isolate verbose work, restrict tools, and enable model routing (cheap Haiku for routine tasks, Opus for complex reasoning). The 12 survivors handle code review, debugging, testing, security, architecture, and data analysis—eliminating the collision and bloat of oversized collections.

• •Only 12 out of 100 tested subagents were worth keeping in real-world use
• •Subagents are context firewalls, not personality roles—their value comes from isolation and tool scoping, not expertise persona
• •Model routing enables cost optimization by assigning cheap models to routine tasks and advanced models to complex reasoning

Generated with AI, which can make mistakes.