Uber Says AI Now Writes 31% of Its Codebase

Uber has revealed that 92% of its developers now use AI agents at least once a month, highlighting the growing role of artificial intelligence in its engineering workflows. 

The figures were shared by company engineers during a presentation at the Pragmatic Summit last month, a conference hosted by engineering management publication The Pragmatic Engineer. 

According to Uber, 31% of code across its codebase is now authored by AI systems, while 11% of pull requests are opened directly by autonomous agents. The company said the tools are part of a broader push to become a “GenAI-powered” organisation and to automate routine engineering tasks.

Internally, Uber has built a layered AI development stack that connects frontier AI models to internal data sources, including source code repositories, documentation, Slack discussions, and engineering tickets. 

Engineers increasingly run multiple AI agents in parallel to generate code, create tests, review pull requests, and handle large-scale migrations. 

The company said that the goal is to eliminate repetitive work such as upgrades, bug fixes and infrastructure migrations, allowing developers to focus more on building product features.

According to a blog post on The Pragmatic Engineer, Uber has built several internal tools to enable large-scale use of AI agents across its engineering teams.

One of the key components is an internal gateway based on the Model Context Protocol (MCP) that serves as a unified interface connecting AI agents to internal services and external tools. The system allows Uber engineers to expose internal endpoints—such as services built with Apache Thrift, Protocol Buffers, or HTTP—as MCP servers with simple configuration changes. 

The gateway also centralises authentication, authorisation, telemetry, and logging, while providing a registry and sandbox environment for developers to discover and experiment with MCP integrations.

Uber has also created an internal no-code platform called Agent Builder, which allows engineers to design and deploy AI agents connected to company data such as source code, documentation, and engineering tickets. The platform includes a visual interface called Agent Studio for building and debugging multi-agent workflows.

To standardise how developers access these tools, Uber built the AIFX CLI, a command-line interface that lets engineers provision AI clients, discover MCP servers, run background agents, and update AI tools across the organisation.

Recently, executives discussed on the company's Q4 2025 earnings call how advances in AI are accelerating the development of autonomous vehicles (AVs).

Chief Executive Officer Dara Khosrowshahi highlighted the role of AI-driven simulation (sim) in training self-driving systems. “Simulation capabilities with larger models, with stronger compute, are getting much, much stronger,” he said.

Khosrowshahi added that combining simulation with real-world driving data is helping speed up development across the industry. “You combine that with advanced simulation capability… and you get to a roadmap, a very fast roadmap to AV readiness for many, many players.”

He also noted that Uber is working with NVIDIA to expand data collection for autonomous driving systems. “We and NVIDIA are teaming up to democratise AV data, real-world data, and provide it to the entire AV ecosystem.”

However, the rapid expansion of AI tools is also increasing infrastructure costs. According to Uber, AI-related spending has risen sixfold since 2024, largely driven by growing token usage across internal developer tools.