Table of Contents
The Evolution of Agentic Engineering at Uber
The landscape of software development is undergoing a seismic shift. At Uber, we have moved beyond simple AI-assisted coding to a strategy centered on agentic AI—an approach that treats artificial intelligence not just as a autocomplete tool, but as an active collaborator capable of managing complex, asynchronous workflows. By enabling developers to offload repetitive tasks to autonomous agents, we have unlocked new levels of engineering velocity and creative potential across the organization.
Key Takeaways
- Shift to Agency: Uber has transitioned from human-led coding to an agentic model where AI handles asynchronous toil, allowing engineers to act as tech leads rather than just code authors.
- Strategic Infrastructure: The adoption of MCP (Model Context Protocol) and custom platforms like Minions and Shephard allows for scalable, secure, and autonomous code management.
- Focus on Toil Reduction: Approximately 70% of AI-driven workloads at Uber are focused on essential but mundane tasks, such as library migrations, documentation, and bug fixes.
- Data-Driven Culture: Success is measured through developer satisfaction and velocity, though the organization is increasingly focused on linking AI impact directly to business outcomes and cost-efficiency.
The Shift from Pair Programming to Peer Programming
In the early days of generative AI, tools like GitHub Copilot functioned primarily as synchronous assistants. While helpful, this represented only a 10-15% bump in diff velocity. Today, we have moved into the era of peer programming. In this model, developers function as directors of their own agentic teams. They define the objective, and the AI agent executes the task asynchronously, coming back to the human only for course correction or high-level strategic input.
AI is enabling people to become superhumans in terms of their productivity and the impact that we can realize for our end users.
This paradigm shift is particularly effective for toil—the necessary but unglamorous maintenance work that keeps a codebase healthy. By automating library migrations, dead code cleanup, and documentation, we allow our engineers to focus on the creative architectural work that actually drives business value.
Building a Robust Agentic Infrastructure
To support this, Uber has built a sophisticated ecosystem of internal tools. Rather than building in isolation, we have integrated our historical Michelangelo platform with modern agentic APIs. A cornerstone of this architecture is the widespread deployment of MCPs (Model Context Protocol), which allows our agents to securely interface with Uber's source code, Jira tickets, and Slack channels.
The Minion Platform
One of our primary innovations is Minion, a formal background agent platform. Unlike generic IDE plugins, Minion runs within our own CI infrastructure, giving it deep access to our monorepos and internal network. It handles everything from prompt improvement—analyzing a user's request to ensure a high success rate—to the automatic creation of pull requests (PRs) and linked Jira tickets.
Optimizing the Code Review Cycle
As the volume of AI-generated code increases, the bottleneck naturally shifts from code authorship to code review. If developers are buried in reviews, productivity gains can quickly vanish. To mitigate this, we have developed several targeted interventions:
- Code Inbox: A unified dashboard that uses smart assignments to route PRs to the most relevant reviewers, accounting for time zones, calendar availability, and historical context.
- U Review: A custom pre-processing engine that filters out low-value "nit" comments and surfaces high-confidence suggestions, ensuring developers only focus on the most impactful feedback.
- Auto-Cover: A dedicated agent for unit test generation that includes a critic engine to validate quality, preventing the influx of low-quality or "change detector" tests into our codebase.
Addressing Organizational and Economic Challenges
Implementing agentic AI is as much a cultural challenge as a technical one. One of the primary hurdles has been adoption friction. Developers who have spent their entire careers operating within an IDE are often hesitant to hand over control to an autonomous agent. We found that the most effective way to overcome this is not through top-down mandates, but by sharing wins—allowing engineers to see the tangible successes of their peers.
The cost of AI is too damn high. We've had to be more responsible about how we use tokens and how we think about what's the right model for the job.
Financial sustainability remains a critical concern. As AI usage has scaled, infrastructure costs have increased significantly. To manage this, we now employ a strategy of model tiering: using high-performance frontier models for planning and reasoning, while utilizing lower-cost, highly optimized models for execution. By treating AI as a managed business resource rather than a magic utility, we ensure that our engineering investments remain aligned with Uber's broader financial objectives.
Conclusion
The journey toward becoming an agentic-powered company has fundamentally changed how we build at Uber. We have learned that the technology is most powerful when it handles the "boring" work that human engineers dislike, and that successful implementation requires a rigorous focus on developer experience and infrastructure security. While the landscape of models and tools will continue to evolve rapidly, our focus remains on providing our engineers with the best possible tools to focus on creative, impactful work that truly scales.
