Claude Code Harness Engineering

The harness is 561k LOC, which is much larger than anticipated.

It seems to show how systems can be built on top of raw LLMs after studying their characteristics.

• Prunes context windows.
• Uses “dreaming” to clean up long-term context.
  • KAIROS prunes after 24 hours of inactivity.
• Is not simply a query-response architecture.
  • It is a loop designed around context-window management.
  • Before the LLM gets any input, a meta-prompt is injected into a state machine.