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

Uber has transformed its software development with an agentic AI-powered factory, leading to a dramatic increase in engineer productivity. The presentation details six key infrastructure components: a unified model gateway with strict PII and safety guardrails, an MCP gateway for streamlined agent tool access and token optimization, agentified dev pods for rapid execution, a managed skills marketplace, a comprehensive context graph, and the Cortana AI assistant. Adam Huda then demonstrates an end-to-end feature development workflow, highlighting a critical shift to inner-loop validation (stopping short of CI) and automated, managed maintenance loops. The ultimate takeaway is that the bottleneck has moved from technical execution to strategic decision-making: "should we build it?" rather than "can we build it?"

August 21, 2026 6 min read

Agentic SDLC at Uber — Uday Kiran Medisetty & Adam Huda, Uber — source: https://www.youtube.com/watch?v=17-YSUHo6Lk

Uday Kiran Medisetty and Adam Huda detail Uber’s journey in building an “agentic AI software factory” that has led to significant productivity gains, including over 70% of pull requests (PRs) originating from local or cloud agents and a doubling of lines of code per engineer year-over-year. This advancement extends beyond coding to every aspect of the engineering lifecycle, including automated toil reduction, with over 250 automated migrations handling 9 million lines of code.

Core Building Blocks of Uber’s Software Factory

Model Gateway: This critical first investment acts as a single, OpenAI/Anthropic-compatible endpoint for all internal and external model calls. Its primary functions include:
- Data Security: Ensures no Personally Identifiable Information (PII) (over 20 types) leaves Uber’s perimeter through a data anonymizer.
- Guardrails: Incorporates five specialized AI safety models to enforce policy, all operating within a strict 100-millisecond latency budget.
- Attribution & Governance: Utilizes Spire for identity and authentication, attributing every request to a specific caller, user, team, and project for comprehensive spend tracking, guardrails, and audit logging. It handles over 100 million requests daily across 800+ projects.

MCP (Model-Compatible Plugin) Gateway & Token Optimization: To address the challenge of agents accessing thousands of internal APIs and various SaaS tools, this gateway provides a unified approach:
- Unified Access: An automated crawler projects internal APIs into MCPs, and the gateway hosts SaaS MCPs (Google, Slack, Jira), handling token exchange for a single entry point.
- Token Efficiency: Initially, an “Omni MCP” allowed discovery and invocation of any MCP via a single installation. Later, a “CLI pattern” was introduced to keep responses out of the context window, resulting in over 40% fleet-wide token savings. A “code mode skill” further hyper-optimizes token-consuming use cases by generating Python scripts on the fly. Uber now has over 1,000 MCP tools.

Dev Pods: Leveraging existing cloud remote development environments, dev pods have been “agentified” to support long-running, isolated, and globally available agent execution.
- Pre-provisioned Environments: Kubernetes “balloon pods” are pre-provisioned with snapshotted repositories and built search indexes, allowing agents to start working in seconds.
- Mega Dev Pods: To accommodate agents and engineers working across multiple repositories and languages, mega dev pods consolidate all repositories into a single environment.

Managed Skills Marketplace: This addresses challenges of skill duplication, discovery, and quality within the organization.
- Centralized Repository: Hosts over 2,500 core and domain-specific skills.
- Quality Assurance: Skills undergo lint checks and automated reviews to ensure baseline quality.
- Discovery & Installation: Offers a single command for discovering and installing plugins, with default skills auto-installed based on engineer personas.
- Continuous Improvement: Collects traces, comments, and continuous evaluations to provide feedback to skill authors.
Context Graph: This crucial knowledge base aims to reduce agent time spent gathering basic context (e.g., service location, dependencies, ownership).
- Unified Knowledge: Replaces 20-30 scattered systems by integrating all Uber’s operational information into one graph with 150 unique node/edge types and 40 million entries.
- Comprehensive Data: Connects mobile apps, backend services, data lakes, design documents, Jira tickets, and incident bugs.
- Performance: Significantly improves token usage, turns, and latency for agents needing context, as demonstrated by queries like analyzing mobility trips.
Cortana (AI Assistant): Uber’s central AI assistant integrates all the above building blocks, providing a unified interface for employees.
- Ubiquitous Access: Available across Slack, CLI, and web, allowing employees to ask questions, invoke skills, and check code.
- Personalization: Employees can personalize Cortana with custom skills and prompts, and integrate it into team Slack channels.
- Adoption: Sees high usage with 300 unique personas created and over 20,000 sessions daily.

End-to-End Feature Development Workflow (Demonstrated by Adam Huda)

Adam Huda illustrates how these building blocks facilitate an end-to-end feature development, from idea to deployment:

Idea Generation & Research: A feature idea (e.g., World Cup pickup locations) starts in Slack. Cortana uses the context graph to research business opportunities, market analysis, and product requirements (e.g., North America rollout).
Design & Prototyping: Cortana assists with Figma designs, generating initial mock-ups (e.g., A/B test variants for button strings). It also suggests code changes, identifying leverageable existing app and backend components, compressing weeks of alignment into a short timeframe.
Code Generation & Inner Loop Validation:
- Minion (Cloud Coding Agent): Uber’s Minion agent, powered by dev pods, works across repositories to generate a draft PR. Crucially, it stops short of pushing to CI to enable earlier validation.
- Shift-Left Validation: The emphasis is on the “inner loop” for validation. This includes static analysis, visual validation (comparing simulator screenshots against Figma specs), and integration testing (bringing up services in a backend staging environment).
- Code Review: Parts of code review are also shifted to the inner loop using smaller, faster models, while the outer loop uses more powerful models for deeper reasoning. Autonomous diffs from Minion include a table of checks performed (e.g., screenshot comparisons) to build human reviewer confidence.
Self-Healing CI & Automated Maintenance:
- Outer Loop: While CI remains, self-healing mechanisms are in place to fix common issues.
- Managed Maintenance Loops: Features and services can be enrolled in managed maintenance skills (e.g., feature flag cleanup). These run on a scheduled loop (e.g., Sundays for CI capacity), control the number of diffs presented to engineers, and use generated diffs and human feedback as data for continuous skill improvement. Incident reviews also feed into the creation of new maintenance skills.

Evolving Bottlenecks

The speakers conclude that with the significant increase in development velocity, bottlenecks have shifted. The focus is now on anticipating and investing in CI capacity, managing the feasible number of experiments, and, most critically, shifting from “can we build it?” to “should we build it?” The ability to build rapidly has made decision-making about what to build the primary constraint.