At Uber, more than 70% of pull requests are attributed to coding agents. Engineers use more than 5K agent skills, and all MCP traffic goes through one gateway with more than 1,000 servers. Between February and August 2026, weekly active users of agentic tools grew 7x and weekly agent requests grew 9.4x and since May the spend has stabilized.

This talk describes the methods behind that. We split agent spend into six terms: users, sessions per user, turns per session, requests per turn, tokens per request, and price per token. For price per token, we built the Uber SWE Benchmark from 1,075 merged pull requests and use it to choose a model and reasoning effort for each workload. For tokens per request, we cap context and match prompt cache lifetimes to the workload. For requests per turn, we use tool search and code-mode batching, and we ground agents in a context graph.  For turns per session, we show engineers a live cost counter as they work.

Key takeaways

• Split cost into terms you can measure, and give each term a lever.
• Choose models per workload from a benchmark built on your own code.
• Most savings come from agent overhead: fewer turns, fewer requests and fewer input tokens per request.