An experiment was conducted to assess the economic benefits of refactoring agent-written code. A large application, primarily in Rust, was built entirely by AI agents. A single data access layer file grew to 17,155 lines.

The refactoring process involved a fresh agent making the same representative change after each refactoring step, allowing for a precise measurement of token consumption. The goal was to demonstrate that refactoring would lower the token cost for future feature implementations.

Results showed a significant reduction in input tokens. Between the baseline and the final refactoring step, input tokens for the same task decreased from 159,564 to 27,360, an 83% saving. This saving is attributed to the agent needing to read less code by successfully identifying smaller, relevant subsets of files. While the overall code size remained constant, the ability to pinpoint necessary sections drastically cut costs. Output tokens remained largely unaffected, but the reduction in input tokens offers substantial economic benefits for ongoing development.