Why forking production AI agents beats building multi-tenant platforms

Short summary

The author argues for forking production AI agents rather than building multi-tenant platforms, because the most expensive thing is not the build but the proven track record. A forked agent inherits its parent's passing eval scores on day one, while a platform refactor invalidates every tenant's numbers simultaneously. With AI agents now handling maintenance and porting fixes between forks, the traditional cost of code duplication has collapsed, making the copy-and-diverge strategy practical.

• •Forking a proven agent preserves its eval track record; a platform refactor voids it for all tenants
• •A medical billing agent took weeks to reach 92% classification accuracy across 150 eval cases — that score attaches to one exact configuration
• •AI-assisted maintenance reduces the traditional cost of keeping forked copies in sync

Generated with AI, which can make mistakes.