Uber's Agentic SDLC Transforms Software Development with AI-Powered Automation

This distributed workforce requires highly efficient and scalable tools to ensure consistent productivity and collaboration.

Uday Kiran Medisetty stated, "we have few thousand engineers across 12 global tech sites."

Agentic AI has revolutionized Uber's development workflow, with more than 70% of pull requests now being generated by local or cloud agents. This automation has led to a remarkable doubling of lines of code per engineer year-over-year, showcasing the profound impact on developer productivity.

This shift has also reduced manual toil through 250 automated migrations, successfully migrating 9 million lines of code automatically. Uday Kiran Medisetty noted, "all of the investments we made in agentic AI have led to more than 70% of our PRs now either by local or cloud agents. And all of this led to twice the number of lines of code per engineer year-over-year."

Uber's current success with agentic AI is rooted in strategic infrastructure investments made over the past six years. The transition to monorepos and the adoption of Bazel provided a robust and scalable foundation essential for accelerating AI integration into the SDLC.

These foundational changes created an environment where AI tools could be effectively deployed and integrated across various projects and teams. Uday Kiran Medisetty explained, "all the investments we made over the last six years on moving to monor repos moving to basil all of that also laid a really solid foundation for us to accelerate this."

Firstly, it strictly prevents Personally Identifiable Information (PII) from leaving Uber's perimeter, safeguarding user data by default.

Secondly, any guardrail added to the gateway must operate within strict latency bounds to avoid impacting performance. Finally, every request processed through the gateway must be attributable per user, per project, and per team, providing transparency and traceability. Uday Kiran Medisetty summarized, "The three things that we wanted to en ensure was no PII ever leaves our parimeter to any of the vendor by default and any guardrail that we add here the latency of that is strictly bounded and every request that goes through this whether it's uh and we need to be able to attribute per user per project and per team."

It begins with identity and authentication via Spire, ensuring secure access and accountability for each request.

Following authentication, a data anonymizer redacts over 20 types of PII, protecting sensitive information. The gateway also incorporates an AI guard with five specialized models that enforce various safety and policy guidelines. Uday Kiran Medisetty specified, "It goes through a series of middleares. The first one is identity and authentication using Spire. Uh we have a data anonymizer that redacts 20 plus PII types. We have a AI guard that has five specialized models that handles various parts of safety and policy that we want to ensure. And all of that runs under 100 milliseconds."

The Uber Model Gateway operates at an impressive scale, currently serving over 800 internal projects. Cumulatively, it manages more than 100 million model requests per day, demonstrating its critical role in Uber's AI infrastructure.

This gateway supports a diverse range of models, including both frontier and open-source solutions, providing flexibility and broad utility across the company's various AI initiatives. Uday Kiran Medisetty stated, "Today we have 800 plus projects internally going through this cumulatively handling more than 100 million model requests per day."

It incorporates a range of middleware functions tailored for engineers, streamlining the integration process.

An automated crawler is a key component, automatically converting internal APIs into MCPs with just a single configuration change, significantly reducing manual effort. The gateway also provides unified token exchange mechanisms for popular SaaS tools like Slack and Jira, enabling seamless agent access.

This centralized approach ensures consistency and efficiency in how agents access and utilize various internal and external resources.

Uday Kiran Medisetty explained, "we have an MCP gateway that handles whole bunch of middleares for for engineers And we have an automated crawler that looks at our internal APIs and projects all of these into MCPS with one single config change."

This proactive approach aims to identify and address issues sooner, reducing the load on CI systems.

By validating features earlier, engineers can avoid the delays and resource consumption associated with CI failures. Uday Kiran Medisetty emphasized, "If we can validate sooner before we push to CI um that would be a big benefit."

For visual validation, a simulator is launched with a specific skill, and screenshots are captured.

These screenshots are then compared against Figma specifications to ensure UI consistency and accuracy. This early validation also includes backend and frontend integration testing, allowing for comprehensive checks before external systems are involved. Uday Kiran Medisetty provided an example, "Things like visual validation. So, we can launch an simulator with a skill, grab a screenshot from the simulator, compare it to the Figma specs."

Uber has implemented self-healing CI, which automatically fixes many common issues encountered during the continuous integration process, thereby reducing manual intervention and speeding up development. This approach optimizes the 'outer loop' of the SDLC.

Code review processes are also being shifted to the inner loop, utilizing different model tiers for efficiency. Smaller, faster models are used for quicker checks in the inner loop, while more powerful models are reserved for deeper analysis in the outer loop, ensuring thoroughness without sacrificing speed. Uday Kiran Medisetty noted, "Self-healing CI is something that we've implemented here where we can fix a lot of the issues that you hit on CI."

This is achieved through enhanced transparency in pull requests.

Validation reports include detailed check tables and screenshots, illustrating the changes made. The system also tracks the self-improvement steps taken by the agent within the diffs, assuring reviewers that the changes have undergone internal vetting. Uday Kiran Medisetty stated, "If this is an autonomous diff coming from minion, we want to give a human reviewer some confidence that this diff has gone through a lot of self-improvement already."

This prevents the proliferation of thousands of unmanaged automation loops across the company.

An example of a managed maintenance skill is feature flag cleanup. These jobs are scheduled during off-peak CI times to optimize resource usage and limit the volume of diffs presented to engineers, particularly on Monday mornings, enhancing overall operational efficiency. Uday Kiran Medisetty explained, "The key thing here is that this is actually a managed loop that you go to, right? We don't want thousands of loops being set up across the company without any bounds."

The agentic system continuously learns and improves by leveraging labeled data generated from daily operations. When an agentic skill creates diffs, the comments, acceptance, or rejection of those diffs provide valuable feedback.

This data is used to refine and enhance the skill itself. Additionally, monthly incident reviews are analyzed to identify opportunities for developing new agentic skills and applying automated maintenance across various services, ensuring a cycle of continuous improvement. Uday Kiran Medisetty highlighted, "When that skill runs and makes those diffs, those diffs will get comments and either get landed or not landed. That's all good label data that we can use to improve the skill itself."

The focus is on anticipating CI capacity requirements and making foundational investments to support future growth.

This shift involves moving beyond assessing technical feasibility to making value-based decisions, ensuring that resources are allocated where they will provide the most impact. Limitations on feasible experiments are also being considered to ensure that efforts are concentrated on high-value initiatives. Uday Kiran Medisetty observed, "We're now putting more strain on our infrastructure. So we're trying to anticipate where our CI capacity needs to be and make the right foundational investments there."