Running a Software Factory Efficiently at Uber Scale (www.uber.com)

At the AI Engineer 2026 conference, Uber unveiled its advancements in using AI agents throughout its software development lifecycle, revealing that over 70% of pull requests are now managed by these agents. They have developed more than 3,600 unique agent skills capable of handling diverse tasks such as code reviews, bug triaging, and automated debugging, achieving over 30,000 executions per day. This shift dramatically reduces human involvement in code-related issues, with active usage of these agents increasing sevenfold in just a few months, indicating a significant transformation in Uber's software development efficiency. This initiative not only enhances productivity but also introduces a sophisticated cost optimization framework. By categorizing agent usage into four layers, Uber can refine cost control and model selection while monitoring key performance indicators associated with each agent's operations. The strategic use of methods such as prompt caching, CLI tool resolution, and context engineering allows Uber to minimize resource consumption significantly. For instance, their new code-mode approach has resulted in over 90% reductions in token usage for bulk workflows, showcasing a robust pathway for other organizations aiming to scale AI integration in their development processes while maintaining cost efficiency.