Uber's 3,600 Agents Handle 70% of Coding as Costs Plunge 52%

AI寒武纪 · wechat · 2026-08-31

More than 70% of coding tasks at Uber are now handled by over 3,600 AI Agents, with weekly requests up 9.4x — yet total costs dropped 52% even as usage skyrocketed. The core methodology includes:

Six-factor cost formula

Spending is broken down into the product of users, sessions, turns, requests, tokens, and unit price, with aggressive optimization applied to the last four factors.

Dynamic model selection

Uber built an internal SWE benchmark and dynamically switches models based on cost, quality, and stability. Sub-agents default to cheap lightweight models, while the main model handles complex reasoning.

Request token optimization

Caching strategy: Interactive session cache TTL adjusted to 1 hour, avoiding frequent context rebuilds.
Tool calls: CLI-ified commands and tool search remove tens of thousands of tokens of redundant MCP Schema from context.
Code patterns: Multi-turn SQL queries and similar operations are bundled into Python scripts, cutting token consumption by over 50%.

AI context graph

An enterprise-scale graph with 24 million nodes lets Agents quickly locate code and data tables, drastically reducing wasted search rounds.

Original post →

Uber's 3,600 Agents Handle 70% of Coding as Costs Plunge 52%

More from coding & agent