Inside Uber’s Developer Infrastructure for AI Agents

In this October 2026 Pragmatic Engineer episode, host Gergely Orosz speaks with Uber distinguished engineer Uday Kiran Medisetty about the company’s development infrastructure and AI agents. Orosz previously worked at Uber; the episode has sponsors. The tools and figures described are the guest’s account of Uber’s practices at the time.

Medisetty joined Uber in 2015 as the company was moving from a small number of services toward a microservice architecture. The hosts describe how rapid growth drove investment in deployment, observability and internal ownership tools.

He explains that separate service repositories made upgrades and builds difficult. Uber moved toward language-specific monorepositories and developed cloud DevPods so engineers could work in prepared environments rather than configuring large builds on laptops.

For AI, the team chose building blocks designed to work across models and agent harnesses. A model gateway presents compatible APIs and centralizes aspects such as policy controls and observability; those controls still require appropriate implementation and review.

Medisetty says an MCP gateway exposes internal tools to agents, using existing service interface definitions to help generate tool specifications. A shared skills repository adds reusable workflows and quality gates across different harnesses.

An internal context graph connects engineering information that had been spread across repositories, ownership records, incidents and other systems. The aim is to help agents and engineers find relevant context with fewer steps; the graph is only as useful as its underlying data and permissions.

The pair discuss product and engineering requirements and the potential for agents and people to work from a shared plan. Medisetty says some parts of that workflow are more mature than others.

Uber’s background coding agent Minion can work from a specification, use cloud development environments and create pull requests across repositories. Medisetty says about 15% of production pull requests went through it at the time; that figure does not establish that the changes were autonomous or independently verified.

The discussion follows generated code into testing and review: consistent build commands help agents, while the team also works on CI repair and review bottlenecks. Dev and test environments remain important parts of the workflow.

Medisetty describes the Edge Gateway and capture-replay tooling as ways to understand requests, dependencies and reliability. Such context can help both engineers and agents investigate problems.

Looking ahead, he frames engineering work as finding the next bottleneck across the entire feature lifecycle. He treats the transition as ongoing and says the surrounding tools and orchestration must evolve alongside models.