Claude Code: Self-Improving Agent Architecture

Most Claude Code setups are static—your agents do the same thing every time. But what if they could learn from their own execution?

In this talk, I’ll demo a self-improving subagent architecture running in Claude Code that uses a plan→build→improve cycle where agents update their own knowledge after every workflow, maintains expertise.yaml files (500-1000 lines per domain) that evolve based on real execution patterns, and shares learnings across domains via a collective knowledge registry so when one agent discovers something, all agents can benefit.

Live demo: I’ll show agents improving themselves in real-time, including the .shared/ folder pattern for cross-domain knowledge propagation.