TL;DR
The Pragmatic Engineer published an interview with Uber Distinguished Engineer Uday Kiran Medisetty about the company’s internal developer infrastructure and its use of AI coding agents. The discussion describes tools for model access, agent configuration, spending controls and code review, while leaving some implementation and productivity details unreported, in an area also explored in Anthropic’s approach to AI.
The Pragmatic Engineer has published an interview with Uber Distinguished Engineer Uday Kiran Medisetty describing how the company’s internal engineering platform supports AI coding agents. The account outlines a model gateway, tools for switching agent interfaces, spending controls and code-review support—systems that matter as Uber adapts its large software operation to AI-assisted development.
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Medisetty, identified by The Pragmatic Engineer as a leader of much of Uber’s work on agentic AI, discussed both tools built before generative AI and newer systems connecting agents with company code, services and engineering knowledge. The interview presents Uber’s approach as an extension of a long-running effort to build infrastructure around the company’s scale and operational needs, a theme also covered in Threlmark’s local-first architecture.
One central component is Model Gateway, a platform team system introduced in 2023 to keep AI applications independent of a single model or harness, according to the report. It supports models from Anthropic and OpenAI, as well as open models Uber runs. The report says security middleware called AI Guard checks requests and responses, while the gateway also supports spend tracking, audit logs and session traces.
Uber’s AIFX layer lets engineers use Claude Code, Codex and OpenCode while carrying settings, skills and plugins across harnesses, the interview says. Engineers receive spending tiers; a manager can approve additional monthly spending when an engineer reaches a limit. A plugin called CostCheck provides suggestions about cost profiles. Separately, Uber’s Code Review tooling helps engineers assess changes after AI review has run and, where relevant, screenshots and videos have been added. The report says engineers still review all pull requests, underscoring the importance of tools that support collaborative work.
How Uber Controls Agent Use
The systems described show how adopting coding agents involves more than choosing a model. At Uber, the reported infrastructure connects agents to engineering work while providing security checks, usage records and cost controls. A shared gateway can also make it easier to change providers or use internally run models without rebuilding each AI application around a single vendor.
The approach may be relevant to other large engineering organizations facing similar decisions: how to give developers access to agents without losing visibility into spending, configuration or code changes. Uber’s account also highlights a continuing human role. AI review and supporting evidence can help prioritize attention, but the report says engineers remain responsible for reviewing every pull request.
Medisetty also described code review as a bottleneck. That makes review tooling a practical part of the AI infrastructure story: generating code is only one stage, and any gains depend in part on how effectively teams can inspect and approve the resulting changes. The interview does not establish that these tools have produced a measured increase in overall engineering productivity.
Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten

Bose Portable Smart Speaker with Alexa and Google Assistant, Black
As an affiliate, we earn on qualifying purchases.
Infrastructure Built for Uber’s Scale
The interview places the AI systems within a history of internal tooling. Medisetty said Uber had two services when he joined in 2015 and thousands by 2019. The report says the service count in 2026 is expected to be about the same as in 2019, despite the business being about 10 times larger than it was then. That 2026 figure is presented as an expectation, not a confirmed later count.
The Pragmatic Engineer says Uber developed tools such as uOwn for service ownership, M3DB for metrics, Cadence for durable workflows and Morpheus for feature flags and experimentation. The report links those choices to needs that available commercial products could not meet at the time. It also describes cities as a core operational concept, reflected in systems for observability, deployments, service sharding and cloud-region placement.
This history helps explain why Uber’s AI setup is presented as a platform rather than a collection of standalone assistants. The company already had internal systems for handling services, deployment and operational data; the newer tools are designed to connect AI development workflows with that environment.
““everything built with AI needs to be model- and harness-agnostic.””
— The Pragmatic Engineer, summarizing Uday Kiran Medisetty’s account
Limits of the Published Account
The report describes the tools and operating principles but does not provide detailed performance data showing how much they have changed deployment speed, defect rates or engineering output. It also does not specify the number of Uber engineers using each agent, the spending limits assigned to tiers, or how managers decide whether to approve additional monthly costs.
The source says Uber reduced token costs by about 50% “the last few months,” referring to earlier coverage, but it does not define the measurement period, comparison baseline or whether that reduction applies across all teams and workloads. The figure should not be read as a company-wide productivity result. The expected 2026 service count is also a projection in the report, not an independently confirmed current total.
Further Details on Adoption and Results
The source does not announce a new product launch or provide a schedule for additional changes. The next useful evidence would be more detail from Uber on how widely engineers use the systems, how spending tiers are set, and how the company evaluates code quality and review time alongside token costs.
For now, the interview offers a view into Uber’s current reported setup: multiple model and harness options, shared infrastructure for monitoring and security, and human review of pull requests. Whether that combination produces sustained gains in engineering speed or quality remains to be shown through disclosed results.
Key Questions
What is the news about Uber’s AI infrastructure?
The Pragmatic Engineer published an interview with Uber engineer Uday Kiran Medisetty explaining how internal tools support AI coding agents, model access, spending oversight and code review.
Which AI coding tools does Uber support?
The report says Uber supports Claude Code, Codex and OpenCode. Its AIFX layer carries settings, skills and plugins between these harnesses.
How does Uber monitor AI usage and costs?
According to the report, Model Gateway supports per-user spend tracking, audit logs and session traces. Engineers have spending tiers, managers may approve additional monthly spend, and CostCheck offers cost-related suggestions.
Does AI review replace human code review at Uber?
No. The report says engineers still review all pull requests. AI review and supporting materials are available before that human review.
Has Uber shown that its AI tools increase productivity?
The interview describes tools and operating practices but does not provide quantified evidence of an overall productivity increase. The reported reduction in token costs lacks a stated comparison baseline and measurement window.
Source: rss
