Agentic SDLC at Uber - Uday Kiran Medisetty & Adam Huda, Uber

Uber has built a managed software factory powered by six key building blocks (model gateway, MCP gateway, agentified dev environments, skills marketplace, co...

By Sean Weldon

Abstract

This paper examines Uber's implementation of an enterprise-scale agentic AI platform that achieves 70% pull request automation while doubling engineering productivity year-over-year. Through six integrated architectural components - model gateway, MCP gateway, agentified development environments, skills marketplace, context graphs, and unified AI assistant - Uber addresses fundamental challenges in enterprise AI adoption: security, governance, knowledge fragmentation, and scalability. The architecture processes over 100 million model requests daily across 800+ internal projects while maintaining sub-100ms latency and comprehensive PII protection. By consolidating scattered systems into unified infrastructure and implementing novel token optimization patterns, Uber demonstrates a replicable framework for autonomous agent deployment across the complete software development lifecycle. These findings provide empirical evidence for enterprise AI strategy and establish quantifiable benchmarks for large-scale agentic system implementation.

1. Introduction

The deployment of autonomous agents in enterprise software development represents a paradigm shift from assistive tooling to autonomous execution. While individual AI coding assistants have achieved widespread adoption, enterprise-scale implementation faces distinct challenges: regulatory compliance, data sovereignty, infrastructure scalability, and organizational knowledge management. This synthesis examines Uber's architectural approach to what they characterize as a "managed software factory" - a comprehensive platform enabling agentic AI systems to operate autonomously across the complete software development lifecycle (SDLC).

Agentic AI, as distinguished from simple code completion or suggestion tools, refers to autonomous systems capable of complex planning, multi-step execution, tool invocation, validation, and self-correction with minimal human intervention. These agents navigate entire codebases, interact with internal APIs, validate changes through multiple mechanisms, and iterate based on feedback. The central thesis posits that enterprise-scale agentic AI requires purpose-built infrastructure addressing six critical domains simultaneously rather than piecemeal tool adoption.

This analysis explores how Uber's integrated architecture achieves measurable outcomes including 70% PR automation, 2x lines of code per engineer year-over-year, 250+ automated migrations handling 9 million lines of code, and 40% fleetwide token savings. The following sections examine the technical foundations, implementation patterns, optimization strategies, and empirical results of this system across a global engineering organization spanning thousands of engineers at 12 technical sites.

2. Background and Related Work

Enterprise AI deployment diverges fundamentally from consumer applications in its requirements for data governance, attribution, auditability, and cost control. Traditional approaches to AI integration typically result in fragmented tooling, inconsistent security policies, redundant effort across teams, and ungoverned model access. The Model Context Protocol (MCP), a standardized interface for AI systems to interact with external tools and data sources, provides the foundation for Uber's unified tool integration strategy.

Prior research in software engineering automation has predominantly focused on isolated tasks within the development lifecycle: code generation, automated testing, or defect detection. Uber's approach synthesizes these capabilities into an integrated platform spanning ideation, design, development, validation, review, deployment, and maintenance phases. The context graph concept adapts knowledge graph architectures from search and recommendation systems to software engineering workflows, consolidating information from 20-30 scattered internal systems into a unified knowledge infrastructure with 150 unique node and edge types representing 40 million entries.

The distinction between inner loop and outer loop development processes proves critical to understanding Uber's validation architecture. The inner loop encompasses rapid, local iteration cycles where engineers test changes before committing code, while the outer loop includes continuous integration (CI), code review, and deployment processes. Shifting validation leftward into the inner loop reduces cycle time, decreases infrastructure load, and enables faster feedback cycles for both human engineers and autonomous agents.

3. Core Analysis

3.1 Security and Governance Infrastructure

The model gateway establishes the foundational security layer for enterprise-scale AI deployment. All model requests route through a single OpenAI/Anthropic-compatible endpoint, enabling centralized policy enforcement while maintaining sub-100ms middleware latency. The architecture implements automatic data anonymization for 20+ PII types, ensuring no personally identifiable information leaves Uber's perimeter to external vendors by default. This approach addresses regulatory compliance requirements while enabling frontier model access.

The gateway's AI guard employs five specialized models for safety and policy enforcement, operating within the strict latency budget. Per-project, per-user, and per-team attribution enables granular spend controls and guardrails across the entire project portfolio. The system captures comprehensive audit logs and session traces, establishing feedback loops for continuous benchmarking and self-improvement. At scale, the gateway handles 800+ internal projects processing over 100 million model requests daily across both frontier and open-source models, demonstrating that centralized governance does not preclude high-throughput operation.

3.2 Tool Integration and Token Optimization

The MCP gateway addresses the dual challenges of tool proliferation and context window consumption. Uber's automated crawler projects thousands of internal APIs into MCP-compatible interfaces requiring only single configuration changes, dramatically reducing integration effort. The gateway additionally hosts and manages SaaS MCPs for external services including Google, Slack, and Jira, handling token exchange and authentication centrally.

Three novel optimization patterns achieve 40% fleetwide token savings. The Omni MCP pattern enables single MCP installation that dynamically discovers and invokes any tools within the gateway, eliminating redundant tool specifications in agent contexts. The CLI pattern optimization prevents response tokens from consuming context windows by structuring tool outputs appropriately. The code mode skill automatically generates Python scripts that hyper-optimize the highest token-consuming MCP operations. These patterns demonstrate that thoughtful architectural design can substantially reduce operational costs at scale, with 1,000+ MCP tools now available through the unified gateway.

3.3 Execution Infrastructure and Knowledge Management

Agentified development environments solve the cold-start problem for autonomous agents through pre-provisioned Kubernetes balloon pods. Agents begin working within seconds with repositories already snapshotted and search indexes pre-built. The mega dev pod consolidates all repositories in a unified environment, enabling agents and engineers to work across codebases without environment switching. This architecture evolved from language-specific dev pods (Go, Java, Android) to support autonomous coding agents and provides global availability with isolation for longer-running agent workloads.

The skills marketplace addresses knowledge reusability through a managed ecosystem of 2,500 skills organized into core and domain-specific categories. Automated lint checks and reviews ensure baseline quality standards for all skills. Default skills auto-install based on engineer personas, eliminating manual discovery and installation overhead. The marketplace processes 20,000+ skill executions daily, with continuous evaluations and trace collection providing feedback to skill authors. This approach transforms ad-hoc automation scripts into managed, quality-controlled components with measurable usage and effectiveness.

3.4 Unified Knowledge and Interface Layers

The context graph consolidates information from 20-30 scattered systems into a single knowledge infrastructure with 150 unique node and edge types capturing 40 million entries. This unified representation spans mobile applications, backend services, data lakes, design documents, Jira tickets, incidents, and bugs. By eliminating the need for multiple skills or MCPs to gather context, the graph substantially reduces token consumption and latency while enabling more predictable agent outcomes. Validation tests demonstrate massive improvements in tokens, turns, and latency when agents query the context graph rather than multiple disparate systems.

Cortana, Uber's unified AI assistant, integrates all building blocks across multiple surfaces including Slack, CLI, and web interfaces. Any company employee can ask questions, look up context, invoke skills, and examine code across codebases. The system supports personalization through custom skills, custom prompts, and team-specific Slack channel integration. Within the first month, 300 unique personas were created, generating 20,000+ sessions daily. This unified interface layer demonstrates that enterprise AI adoption extends beyond engineering teams when infrastructure properly abstracts complexity.

4. Technical Insights

Uber's end-to-end feature development workflow illustrates how architectural components integrate across the complete SDLC. During ideation, Cortana leverages the context graph to evaluate business opportunities and product requirements. In the design phase, the assistant helps with Figma designs and creates A/B test variants. The Minion agent, Uber's cloud coding agent, handles backend and frontend changes across multiple repositories during development.

The validation architecture implements a shifted-left strategy with inner loop validation encompassing static analysis, visual validation via simulator screenshots, and backend integration testing before pushing to CI. Code review employs smaller, faster models in the inner loop while reserving powerful reasoning models for deeper outer loop review. Self-healing CI autonomously fixes issues detected in the continuous integration pipeline. Maintenance phase automation runs managed skills on controlled schedules (e.g., Sunday) with controlled diff landing and feedback loops for continuous skill improvement.

This architecture reveals critical implementation considerations. CI capacity constraints require anticipatory infrastructure investments as agent-generated code volume increases. Limited experiment capacity creates validation bottlenecks despite increased development velocity. The decision-making paradigm shifts from "can we build" to "should we build" as technical capability expands faster than strategic prioritization. These bottlenecks suggest that infrastructure and process scalability, rather than AI capability, increasingly constrain enterprise automation efforts.

5. Discussion

Uber's implementation provides empirical evidence that enterprise-scale agentic AI requires integrated infrastructure rather than point solutions. The 70% PR automation rate and 2x productivity improvement demonstrate that autonomous agents can handle substantial portions of software development when provided with appropriate security, tooling, execution environments, reusable knowledge, unified context, and accessible interfaces. The 250+ automated migrations processing 9 million lines of code illustrate capabilities extending beyond feature development to large-scale refactoring and maintenance.

The token optimization strategies achieving 40% fleetwide savings reveal that operational costs at scale demand architectural attention beyond model selection. The Omni MCP pattern, CLI pattern optimization, and code mode skill demonstrate that thoughtful interface design substantially impacts economics. Similarly, the context graph's consolidation of scattered systems addresses not only agent effectiveness but also cost efficiency by reducing redundant context gathering operations.

Several areas warrant further investigation. The bottleneck shift from development to CI capacity and experiment validation suggests that infrastructure and process scalability may limit automation gains more than AI capability. The decision-making evolution from technical feasibility to strategic prioritization raises questions about organizational readiness for dramatically increased development velocity. The 300 unique personas created in Cortana's first month indicate demand for AI assistance beyond traditional engineering roles, suggesting broader organizational transformation potential.

6. Conclusion

This analysis demonstrates that enterprise-scale agentic AI deployment requires comprehensive architectural investment across six integrated domains: secure model access, unified tool integration, scalable execution infrastructure, managed knowledge components, consolidated context, and accessible interfaces. Uber's implementation achieves measurable outcomes including 70% PR automation, 2x engineering productivity, and 40% token savings while processing 100+ million daily model requests across 800+ projects.

The practical takeaways establish a replicable framework for organizations pursuing autonomous agent deployment. Security and governance must be centralized to ensure compliance without sacrificing latency. Tool integration requires thoughtful optimization patterns to manage context window consumption at scale. Execution infrastructure must eliminate cold-start delays through pre-provisioning. Knowledge management demands quality controls and reusability mechanisms. Context consolidation substantially improves agent effectiveness and cost efficiency. User interfaces must abstract complexity to enable organization-wide adoption.

Future work should examine the organizational and process adaptations required as technical constraints diminish relative to strategic and infrastructure bottlenecks. As agentic AI capabilities continue advancing, the limiting factors for enterprise automation may increasingly shift from "what can agents do" to "how should organizations restructure workflows, decision-making, and infrastructure to leverage dramatically increased development velocity."

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub