Uber says years of developer platforms made AI agents far easier to adopt

Uber says years of unglamorous work on developer platforms made it much easier to put AI coding agents to work. That was the account of Uday Kiran Medisetty, a Distinguished Engineer who leads much of Uber's agentic AI work. He spoke on The Pragmatic Engineer in an episode published on 14 October 2026.

What was said

Medisetty described three pieces. The first is DevPod. Around 2017-18, Uber moved its services into one big code repository per language. Setting those up on a laptop was painful, so Uber built cloud developer boxes with builds already cached. Agents now use the same boxes. Because build and test commands are standardized, one top-level instruction file tells an agent how to work in any module.

The second is a single registry of service interface descriptions, written in Thrift or protobuf. Uber used it to auto-generate MCP specs for its services. MCP is a standard way of handing tools to AI agents.

The third is Slate. An engineer deploys a branch into a private test instance and routes a test account to it. The rest of the request path uses normal staging. Uber then taught agents to deploy Slate themselves. Medisetty said that "if we did not have like system like slate it could have been really hard for agents to self-validate". Of the older platforms, he said "you could use all the things that we built making the life easier for engineers and it makes it easy for our agents as well".

Why it matters

The other side

This is one company's story, and a very large one. One speaker said Uber built much before 2022, so adding AI infrastructure was fairly easy. That speaker conceded it was still work. Later, a speaker on the show said the companies gaining most from AI are those that had developer platform teams before it. That is an opinion voiced on the show. The episode did not test firms without such teams.

Medisetty was candid about limits. He said a first version of an agent skill is easy to make, but making it good across many use cases is hard. A speaker on the show said engineers cannot review every pull request with the rigor of the past. Medisetty said Uber even had to add an internal "code red" for its CI system. The excerpts do not say why. Old platforms give agents a base, but they do not settle who checks the output.