Uber's Minion agent writes 15% of production PRs, all still human-reviewed

Medisetty's account traces a deliberate sequence: Uber wrapped every coding harness in an internal layer called AIFX so settings, skills and plugins carry over when engineers switch between tools like Claude Code and Codex. It then built a single MCP gateway in around April 2025, auto-generating tool specifications from its Thrift and Proto IDL registry. That groundwork matters because Medisetty said skills run roughly 30,000 times a day and all use the gateway, keeping access uniform.

A second thread is context. Uber's context graph, built over recent months, links more than 100 node types and over 130 edge types, connecting people, assets, design docs, incidents and Jira. Medisetty said assistants use it as a default scaffold, narrowing the question area before reading deeply, which cut round trips and lowered cost and latency. That claim is his own characterization, not an independently measured figure.

On work itself, Medisetty described a capability overhang: models can already do far more than current orchestration and tooling let them apply. He said he personally writes far more PRs than last year, and that engineering work increasingly means building the context agents need, making context engineering and well-specified constraints the core skill.