Scaling Agentic SDLC at Uber
AI Engineergo watch the original →
the gist
Uber scaled its agentic software development by building a centralized infrastructure layer that handles PII redaction, context graph indexing, and tool orchestration, resulting in 70% of PRs now being agent-generated.
Centralized Infrastructure for Agentic Workflows
Uber implemented a unified model gateway that acts as the single entry point for all internal AI requests. This gateway enforces security and compliance by running Spire-based identity authentication, redacting over 20 types of PII, and executing five specialized safety models, all within a strict 100-millisecond latency budget. This infrastructure supports 800 projects and handles over 100 million requests daily, providing full attribution per user, team, and project.
Reducing Token Tax and Context Overhead
To prevent agents from drowning in irrelevant data, Uber built an MCP (Model Context Protocol) gateway that crawls internal APIs and projects them into a CLI-based format. This approach keeps responses out of the LLM context window, contributing to a 40% reduction in fleetwide token usage. Furthermore, a centralized context graph containing 150 node and edge types and 40 million entries replaces the need for agents to crawl 20 to 30 disparate systems for basic ownership and dependency facts.
The Agentic Software Factory
Development environments are now pre-provisioned as Kubernetes balloon pods, allowing agents to start work in seconds with snapshotted repositories and indexed search. A managed skills marketplace hosts 2,500 entries, each subjected to automated linting and review, which facilitates over 20,000 executions daily. For end-to-end feature development, agents like 'Minion' generate draft PRs that stop short of CI, allowing for inner-loop validation where visual outputs are compared against Figma specs before any code reaches the build queue.