Uber Reveals Strategy for Operating AI Agent-Based 'Software Factory'... Cutting Costs by 34%

As AI tools are integrated into every stage of the software development lifecycle, Uber has revealed its strategy and cost optimization methods for efficiently operating a 'software factory.' Currently, over 70% of all Pull Requests at Uber are generated by local or cloud agents, with an average of more than 30,000 agent skill executions per day.

Four-Tier Agent Usage Hierarchy

Uber classifies AI usage into four tiers based on the degree of specialization. Higher tiers offer greater control over cost, quality, and model selection. This structure allows Uber to decompose and optimize the costs of tasks performed by agents.

Cost Decomposition and Optimization Metrics

Total spending is decomposed and measured across six items, including adoption rate, engagement, and agent self-workload. Specifically, when compared against a fixed model baseline regardless of model upgrades, the model request cost per 1,000 requests decreased by approximately 34% from the peak, and the cost per session decreased by 52% from the June peak. These results were achieved through input token optimization, reduction of unnecessary turns, and utilization of prompt caching.

Key Management Metrics

Uber tracks the following metrics on weekly and monthly bases to formulate short- and long-term plans.

• Portfolio Level: Total attributed cost, unique user count, cost share by tool
• Unit Economics: Cost per user, tokens per request, prompt cache hit rate
• Model Economics: Cost and request share by model, cost change per token
• Managed Agent Performance: Cost per merged PR, revert rate, MTTR, and other quality signals