Uday Kiran Medisetty describes six building blocks behind Uber’s agentic development workflow: a model gateway, an MCP gateway, cloud development environments, a managed skills marketplace, a context graph and an internal assistant. The gateways centralize authentication, privacy controls, request attribution and tool access, while tool discovery, CLI access and generated scripts reduce context overhead. Pre-provisioned environments let agents work across repositories with code snapshots and search indexes ready.
Uday Kiran Medisetty explains how Uber manages skills through automated checks, simpler discovery and installation, and feedback from execution traces and evaluations. A shared context graph connects code, ownership, dependencies, design documents and operational information so agents can retrieve context without repeatedly traversing separate systems. These capabilities feed an assistant available through Slack, the web and a command line; the components are at different stages of maturity and rollout.
Adam Huda walks through a stadium pickup feature, from business research and requirements to design variants and implementation by a cloud coding agent. The workflow initially stops at a draft pull request so agents can check the feature before consuming CI capacity. Earlier validation includes static analysis, simulator screenshots compared with design specifications, frontend and backend integration, and faster code review, with deeper review and repair continuing in CI. A record of checks and screenshots gives human reviewers evidence of the work already performed.
Adam Huda describes managed maintenance loops that clean up feature flags and improve services using feedback from reviews, accepted changes and incidents. Central scheduling aims to use available CI capacity without flooding engineers with pull requests. As agents increase code output, the remaining constraints include infrastructure capacity, how many experiments an organization can run, and deciding which features are worth building.
Watch on YouTube
