Uber's Minion agent generates 15% of production pull requests

Uber's background coding agent, Minion, currently generates 15% of all production pull requests, a figure that Udaykira Mendicetti confirmed in recent disclosures. The latest iteration, Minion v2, operates through web and Slack interfaces within a 'Mega DevPod' environment that grants access to all repositories for cross-repository code generation. Engineers can steer the agent directly within Slack threads, enabling immediate bug fixes without waiting for backlog clearance. To manage the resulting volume, Uber deployed 'CodeInbox' to aggregate review signals, including AI code review results, sanity test screenshots, and risk scores. The company also utilizes 'Slate' for end-to-end testing, which routes specific test account UIDs to deployed branches across the service graph using Jaeger tracing baggage, while 'Capri' captures and replays production traffic to validate fixes against specific dimensions like city or payment type. Additional tools such as 'Fox' automate weekly maintenance tasks like dead code cleanup, and 'Shepard' manages large-scale library migrations. Uber reports that the percentage of feature PRs is increasing while maintenance PRs decrease, indicating improved feature velocity. Mendicetti argues that production-ready code requires more than an AI harness; it depends on retrofitting AI onto battle-tested developer infrastructure to enable self-validation. Uber’s current capabilities rely heavily on infrastructure built before the LLM era, including the transition to language-specific monorepos around 2017-2018 and the introduction of DevPod in 2017. These pods, warmed up on Kubernetes clusters, launch in seconds and provide standardized tools that allow agents to execute identical build and test commands as human engineers. The company also developed a 'Context Graph' with over 100 node types and 130 edge types to interconnect people, assets, and incidents, serving as the default grounding for its internal AI assistant, Cortana. An 'MCP Gateway,' built starting in April 2025, auto-generates Model Context Protocol specs from existing Thrift and Protobuf registries, exposing internal services as MCP servers. A 'Model Gateway' built on the Michelangelo ML platform routes all model requests through a single entry point, enforcing security policies and tracking spend. Mendicetti emphasized that investing in harness- and model-agnostic building blocks early allowed Uber to port across changing AI landscapes and leverage existing API infrastructure.