The Infrastructure of an Agentic Factory

Uber has transitioned to an agent-first software development lifecycle (SDLC) by implementing six core infrastructure building blocks. This system currently handles over 100 million model requests daily across 800 projects, resulting in a 2x increase in lines of code per engineer year-over-year.

  • Model Gateway: A centralized, OpenAI/Anthropic-compatible endpoint that enforces security and compliance. It handles Spire-based identity, redacts 20+ PII types, and runs five specialized safety models—all within a strict 100ms latency budget.
  • MCP Gateway & Token Optimization: To prevent "token tax," Uber uses an automated crawler to project internal APIs into Model Context Protocol (MCP) servers. By projecting these into a CLI pattern and using a single "Omni MCP" entry point, they reduced fleetwide token consumption by 40%.
  • Agentified Dev Pods: Pre-provisioned Kubernetes "balloon pods" provide isolated, snapshotted environments that allow agents to begin work in seconds, supporting cross-repository tasks through a unified "mega dev pod" architecture.
  • Managed Skills Marketplace: A curated repository of 2,500 skills that undergo automated linting and review. This ensures baseline quality and allows agents to automatically discover and execute tasks, currently performing 20,000+ executions daily.
  • Context Graph: A massive knowledge graph containing 40 million entries across 150 node/edge types. It maps dependencies, ownership, design docs, and incident history, allowing agents to retrieve accurate context without crawling dozens of disparate systems.
  • Cortana Assistant: A unified interface across Slack, CLI, and web that integrates the above components, allowing employees to build, research, and collaborate using personalized agent personas.

Shifting Validation to the Inner Loop

Uber’s approach to the SDLC emphasizes "shifting left" by moving validation out of the expensive CI outer loop and into the agentic inner loop.

  • Visual & Integration Validation: Before pushing to CI, agents use skills to launch simulators, capture screenshots, and compare them against Figma design specs. They also validate backend-frontend integration in staging environments.
  • Self-Healing CI: When code does reach the outer loop, Uber employs self-healing mechanisms to automatically fix common CI failures.
  • Managed Maintenance Loops: Rather than ad-hoc scripts, maintenance tasks (like feature flag cleanup) are enrolled in managed loops. These are scheduled during off-peak hours to optimize CI capacity and prevent overwhelming engineers with excessive diffs on Monday mornings.

The New Engineering Bottleneck

By automating the mechanics of coding, Uber has fundamentally shifted the primary constraint of software development. The bottleneck is no longer the technical ability to implement a feature, but the strategic decision-making process regarding whether a feature should be built at all. The focus has moved toward using AI to assist in business research, product requirement definition, and impact analysis before a single line of production code is written.