Descriptions:
Uber’s developer platform organization, represented by engineering VP Unshu and principal engineer Tai, details how the company has moved from AI-assisted coding to fully agentic engineering workflows — and what that transition has required technically, organizationally, and culturally. The talk is grounded in real metrics: GitHub Copilot’s tab completion and chat features produced a 10–15% increase in diff velocity, while the shift to asynchronous background agents has produced substantially larger gains for toil-heavy work like dead code cleanup, documentation generation, and library migrations.
The technical centerpiece is Minions, Uber’s internal agentic task platform built on top of Claude Code. Minions provides a web interface where engineers submit tasks against any of Uber’s monorepos, choose from curated prompt templates, and receive Slack notifications with links to completed pull requests — typically within minutes. A built-in prompt quality scorer flags weak prompts before the agent runs and suggests improvements, directly improving task success rates. The demo shows a bug fix task completing in seven minutes, resulting in a PR co-authored by the Minion bot with a full test plan.
Beyond the technical implementation, the talk candidly addresses the non-technical barriers to agentic adoption at scale: measurement challenges (how do you attribute productivity gains when agents run asynchronously?), cost management, and the cultural shift required for engineers to act as tech leads directing agents rather than writing code directly. Uber’s framing of this as a move from “pair programming” to “peer programming” offers a concrete mental model for engineering organizations at similar inflection points.
📺 Source: The Pragmatic Engineer · Published March 10, 2026
🏷️ Format: Workflow Case Study
