- A speaker on The Pragmatic Engineer said the coding harness is the easy part. The hard part is everything around it.
- Uber's Uday Kiran Medisetty said a first skill is quick to write but hard to make good, and tool setup can crowd the agent's context.
A speaker on The Pragmatic Engineer argued that the coding agent is the easy part of AI-assisted engineering. The hard part is the system around it: shared skills, tool links and quality checks. The point comes from the 14 October 2026 episode "Uber's AI infra, from the inside." Uday Kiran Medisetty supplied the detail. He is a Distinguished Engineer who leads much of Uber's agentic AI work.
What was said
The episode's introduction set the theme. Getting code ready for production takes far more than running an AI coding tool, a speaker said. In that speaker's words, "The harness is the easy part." The harness is the program that runs the model and lets it work on code.
Medisetty described what the hard part looks like at Uber. Skills are saved instructions that teach an agent a task. Most live in one shared repository. Central platform teams write the core ones. Individual teams add their own. He put recent use at around 30,000 skill runs a day.
A first version of a skill is quick to write, he said. The trouble comes later. In his words, "it's very hard to make it really good across all kinds of use cases."
Upkeep has a technical price too. Each skill can name the tool servers it needs. Medisetty said that installing two plugins can leave you with six of them. That crowds the agent's context, the text it must hold in view. Two engineers, using AI, rewrote the skills to call tools through the command line. He said this used fewer tokens. Tokens are the units of text a model reads.
His team is also building a way to collect traces from agent sessions. Those traces would feed draft patches that improve the skills.
Why it matters
Our reading: the work moves from writing code to judging and tending the setup that produces it. A first version of a skill costs little. Keeping it good across many cases is ongoing work. It needs an owner, tests and a way to see failures.
Tool setup is a running cost too. Everything the agent carries takes up room in its context. Managers and buyers can ask three questions. Who owns each skill? How is its quality measured? What does each tool add to the agent's load?
The other side
Uber may be an unusual case. A speaker on the show noted a pattern: Uber built tools over many years before AI arrived. Another speaker said the companies gaining most from AI had developer platform teams before it. A smaller team may lack that base.
The excerpts also leave questions open. The trace-based skill improvement was still being built, so we cannot say it works. The 30,000 figure measures use, not quality. Medisetty also said there is a lot of discussion about skill quality and evals.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- The Pragmatic Engineer: Uber’s AI infra, from the inside (2026-10-14)
