How Uber Built AI Agents That Save 21,000 Developer Hours with LangGraph | LangChain Interrupt

Summary

In this presentation, Matasanis and Sorup Sherhhati from Uber discuss how they built AI-powered developer tools using Langraph to streamline development processes. With a massive codebase and 5,000 developers, Uber's developer platform focuses on automating tasks like code validation, test generation, and security checks. They showcase tools like Validator for automatic code fixes and Autocover for generating high-quality tests. The presentation highlights key insights on agent composition, the use of deterministic agents, and the impact of AI in scaling development efforts, demonstrating substantial improvements in productivity and code quality at Uber.

Takeaways

• 😀 Uber serves 33 million trips a day across 15,000 cities, powered by a massive codebase with hundreds of millions of lines of code, which the developer platform team supports.
• 😀 Uber's AI dev tool strategy focuses on three pillars: building products that improve developer workflow, creating crosscutting primitives, and ensuring intentional tech transfer to reuse tools across various projects.
• 😀 Langraph is a central tool in Uber's AI development strategy, enabling reusable agents to solve problems in an agentic manner and integrating with Uber systems through a framework called Lang effect.
• 😀 The 'Validator' tool is an IDE plugin that flags best practices violations and security issues in code automatically, offering precomputed fixes and agent-driven solutions.
• 😀 The 'Autocover' tool helps engineers quickly generate high-quality test coverage, automating the creation of business case tests and performing mutation testing to ensure test reliability.
• 😀 Uber leverages Langraph agents to create a sophisticated test generation process, where multiple tests are executed in parallel to increase speed and coverage, saving developer time.
• 😀 Uber's 'Uber Assistant Builder' allows internal teams to build custom GPT-powered bots that are steeped in Uber-specific knowledge, such as security best practices and architecture queries.
• 😀 'Picasso' is Uber's internal workflow management platform, with a conversational AI (Genie) that understands workflow automation and product truths, enhancing decision-making across teams.
• 😀 'U Review' helps reinforce code quality by flagging anti-patterns and suggestions during code review, improving the development process before PRs are merged.
• 😀 Key learnings include the importance of building highly capable domain expert agents, composing deterministic sub-agents for reliability, and scaling efforts by reusing agents across multiple applications to boost developer productivity.

Q & A

What is the core focus of the AI developer tools at Uber?

The core focus of the AI developer tools at Uber is to improve developer workflows by automating repetitive tasks, enhancing productivity, and reducing toil for developers. This includes tools like code validation, test generation, and workflow automation.

How does Uber’s AI strategy align with developer needs?

Uber’s AI strategy is built on three pillars: products that directly improve developer workflow, foundational AI technologies that enable faster development, and intentional tech transfer that makes tools reusable across different problems and teams.

What is Langraph, and how does it play a role in Uber's AI developer tools?

Langraph is a framework that Uber built, which wraps around Langraph and Lang Chain to make them work more efficiently with Uber’s systems. It was designed to create reusable nodes and build agentic systems that can be leveraged across multiple tools and products.

Can you explain how the 'Validator' tool works at Uber?

The 'Validator' tool in Uber is an integrated development environment (IDE) tool that automatically detects best practices violations and security issues in code. It provides feedback to engineers, suggesting precomputed fixes or allowing them to request fixes via an agent assistant.

What is the role of 'Autocover' in Uber’s developer toolset?

'Autocover' is a tool designed to help engineers quickly generate high-quality tests for their code. It uses domain-specific agents to build, test, and validate coverage for new business cases and mutations, saving time for developers by automating the test generation process.

How does Uber's AI-driven test generation process work?

The test generation process involves analyzing source files and adding new targets to the build system. Once a target is identified, it runs coverage checks and generates tests in real-time. The tool continuously refines the tests, removing redundant ones and adding necessary tests, ensuring comprehensive coverage.

What have been the results of using these AI tools in terms of developer productivity?

These AI tools have significantly improved developer productivity by automating tasks like test generation and code validation. The tools have helped raise platform coverage by about 10%, saving approximately 21,000 developer hours and generating thousands of tests monthly.

What strategic benefits does encapsulation bring to Uber's AI development?

Encapsulation boosts collaboration by creating well-defined abstractions, enabling developers to focus on specific tasks. This approach allows Uber to scale development horizontally, improve productivity, and tackle more complex problems without creating operational bottlenecks.

How does Uber use graphs to enhance AI agent collaboration?

Graphs are used to model interactions between different agents and workflows. By capturing how developers naturally interact with the system, Uber can identify process bottlenecks and improve both AI workloads and the experience for developers not directly using the AI tools.

What are some additional tools Uber has developed using the Langraph framework?

In addition to 'Validator' and 'Autocover,' Uber has developed tools like the 'Uber Assistant Builder,' which creates custom chatbots for various organizational needs, and 'Picasso,' an internal workflow management platform with AI assistance, as well as 'U Review,' a tool for reinforcing quality checks during code review.