How to Build an AI-Native Software Factory: Uber’s Shift from Laptops to Managed Agents

The article outlines how to build an AI-native software factory, using Uber’s experience as a primary case study. Uber’s core strategic shift involves moving from developers running agents on local laptops to a centralized platform that offers complete control over model routing, harnesses, and spend. This transition addresses key limitations of the laptop model: the inability to run agents without a human present, the lack of shared security rules and skills, and the invisibility of total AI costs until invoicing.

Uber’s AI costs increased 6x since 2024, with the annual budget exhausted in just four months by June, prompting a $1,500 monthly cap per employee per tool. To manage this, Uber implemented a model gateway that tracks every request by user, project, and team, and introduced metrics like 'cost per unit of work' (e.g., per merged pull request). The platform is structured around six building blocks: a model gateway, an MCP gateway for tool access, cloud development environments (DevPods), a skills marketplace, a context graph linking 24 million nodes, and a front-door assistant (Cortana).

The article emphasizes that companies should not build proprietary coding agents but rather purchase existing ones (like Claude Code, Codex, or OpenCode) and focus on building the surrounding infrastructure. It recommends starting with a 'thinnest viable platform'—buying the pipes (gateways, sandboxes) and building only the specific context and workflows unique to the company. The piece also highlights the importance of measuring merged work rather than code written, citing Uber’s 52% reduction in cost per session after implementing visibility tools, and Intercom’s success in tripling merged pull requests per employee.