Uber's AI-Native Developer Infrastructure Strategy
An executive analysis of how Uber leverages a decade of platform engineering to deploy agentic AI. The company uses unified gateways, context graphs, and automated validation to shift engineering focus from code volume to feature velocity, ensuring scalable and secure AI integration across thousands of services.
Strategic Shift to Agentic Infrastructure
Uber’s engineering leadership has fundamentally restructured its development lifecycle to accommodate agentic AI, leveraging a decade of platform engineering investments. The core strategic insight is that AI adoption is not merely about deploying new models, but about retrofitting existing, battle-tested infrastructure to support autonomous agents. By standardizing on monorepos and strong API contracts, Uber can auto-generate Model Context Protocol (MCP) specifications, allowing agents to interact with internal tools without custom integration work. This approach minimizes friction and ensures that AI capabilities scale alongside the company’s complex, multi-service architecture.
Governance and Cost Control
A critical component of Uber’s strategy is the implementation of unified gateways for both model access and tool invocation. The Model Gateway and MCP Gateway serve as single entry points for all AI interactions, enabling centralized security policies, audit logging, and cost attribution. This architecture allows the company to enforce spend tiers and monitor usage across all teams, preventing the uncontrolled consumption of tokens that often plagues early AI adopters. By abstracting the underlying models and tools, Uber maintains vendor neutrality, ensuring that rapid shifts in the AI landscape do not disrupt internal workflows.
Redefining Productivity Metrics
Uber has moved beyond traditional metrics like code volume to focus on feature velocity. By tagging pull requests with complexity levels and business categories, the company tracks the time from product requirements to production deployment. This shift ensures that the increased throughput provided by AI agents translates into tangible business value, such as new features, rather than just increased maintenance tasks. The engineering culture is evolving to prioritize context engineering, where developers focus on defining precise specs and verification loops for agents rather than writing every line of code manually.
Operational Impact
The integration of systems like Slate for environment isolation and Capture Replay for production traffic simulation allows agents to self-validate their work. This reduces the burden on human reviewers, who can now focus on high-risk changes based on automated risk scores. As Uber continues to refine these tools, the role of the software engineer is transitioning from a code producer to an orchestrator of autonomous systems, requiring new skills in multitasking and quality assurance.
Key insights
-
Pre-existing platform infrastructure, such as standardized APIs and monorepos, is the primary enabler for successful agentic AI adoption. Without these foundations, integrating agents into complex systems becomes prohibitively difficult and error-prone.
Impact: Companies with strong platform engineering teams can deploy AI agents faster and more securely than those starting from scratch, creating a significant competitive advantage in operational efficiency.
-
Centralized gateways for model and tool access are essential for governing AI usage at scale. These gateways enable strict cost control, security auditing, and vendor-agnostic switching, preventing shadow AI and uncontrolled spend.
Impact: Organizations that implement unified AI gateways can reduce operational costs and mitigate security risks, ensuring that AI adoption remains compliant and financially sustainable.
-
Context graphs that connect people, code, and operational data significantly improve agent performance by reducing the number of turns required for accurate responses. This grounding mechanism lowers inference costs and improves the reliability of AI-generated insights.
Impact: Investing in knowledge graph infrastructure can lead to substantial reductions in AI operational costs and improved accuracy, making agentic systems more viable for complex enterprise tasks.
-
The definition of engineering productivity is shifting from code volume to feature velocity. By tracking the time from product requirements to production, companies can ensure that AI-driven throughput translates into actual business value rather than just increased maintenance.
Impact: Adopting feature-velocity metrics helps organizations align AI investments with business outcomes, ensuring that increased developer productivity results in faster time-to-market for new features.
-
The role of the software engineer is evolving from code producer to context engineer. Success now depends on the ability to define precise specs, constraints, and verification loops for autonomous agents, requiring new skills in orchestration and quality assurance.
Impact: Companies must invest in training engineers for context engineering and agent orchestration to fully realize the productivity gains from AI, as traditional coding skills become less differentiating.
Action items
-
Audit existing API documentation and code repositories to identify opportunities for auto-generating MCP specifications. This will enable agents to interact with internal tools without custom integration work.
Impact: Reducing integration friction will accelerate the deployment of agentic workflows and lower the total cost of ownership for AI initiatives.
-
Implement a centralized gateway for all AI model and tool interactions to enforce security policies, track costs, and enable vendor-agnostic switching. This should include strict spend tiers and audit logging.
Impact: Centralized governance will prevent shadow AI usage, reduce operational costs, and ensure compliance with security and data privacy regulations.
-
Develop a context graph that connects key organizational data points, such as code ownership, incident history, and design documents. Use this graph to ground AI agents and reduce the number of turns required for accurate responses.
Impact: Improved grounding will lead to higher accuracy and lower inference costs, making agentic systems more reliable and cost-effective for complex tasks.
-
Shift productivity metrics from code volume to feature velocity by tracking the time from product requirements to production deployment. Tag pull requests with business categories to align AI-driven throughput with business value.
Impact: Focusing on feature velocity will ensure that AI investments translate into tangible business outcomes, such as faster time-to-market and improved customer satisfaction.
-
Train engineers on context engineering and agent orchestration, emphasizing the importance of defining precise specs, constraints, and verification loops. Provide tools and frameworks to support these new workflows.
Impact: Upskilling the workforce will maximize the productivity gains from AI adoption and ensure that engineers can effectively manage and verify autonomous agent outputs.
Quotes
“The harness is the easy part. What's hard is all the things around them.”
“We wanted to invest in the right set of building blocks that are hardness and model agnostic so that no matter any of these things change we can easily port across different models different harnesses”
“The job of the engineering is you know in some sense how do you build the right context for the agents to produce reliable output.”