TL;DR
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
The Pragmatic Engineer published an interview with Uber Distinguished Engineer Uday Kiran Medisetty about the company’s developer infrastructure and its use of AI agents. The discussion describes systems for routing models, managing agent tools and costs, and supporting human code review; it does not provide independent productivity measurements or a full account of deployment outcomes.
The Pragmatic Engineer has published an interview with Uday Kiran Medisetty, a long-serving Uber engineer who leads work on agentic AI, describing how the company connects coding agents to its engineering systems. The account outlines Uber’s model gateway, agent tooling, spending controls and code-review process, offering a view of how a large technology company is adapting developer infrastructure for AI-assisted work.
Medisetty describes a stack built partly on systems Uber developed before generative AI became widely available. The interview says Uber now supports multiple coding harnesses, including Claude Code, Codex and OpenCode. A wrapper called AIFX carries settings, skills and plugins between supported harnesses, while the Model Gateway lets teams switch among vendor models and models Uber runs itself.
The gateway also provides a point for security checks and operational monitoring. According to the report, AI Guard, run by Uber’s security team, checks requests and responses and measures latency. The platform records spending by user and keeps audit logs and session traces. These controls are intended to give Uber visibility across different models and tools, though the source does not publish detailed performance or security results.
Uber uses spending tiers for engineers, the report says. When an engineer reaches a monthly allocation, a manager can approve additional spend; the policy is not described as a fixed universal cap. A plugin called CostCheck provides suggestions across sessions. The article reports that Uber has reduced token costs by about 50% in recent months, citing earlier coverage, but does not give the measurement period, baseline or methodology in the material provided.
Human engineers still review pull requests. The report says an internal Code Review tool helps reviewers decide how much attention a change needs; AI review, relevant screenshots or videos, and risk scores can be available before a person reviews the code. Medisetty describes code review as a bottleneck, but the article does not quantify how much the tools have reduced review time.
How Uber Governs Coding Agents
The interview illustrates that deploying AI coding tools at a company with many engineering teams is not just a matter of giving employees access to a model. Uber’s approach, as described, places model choice, security checks, spending visibility and human review within shared infrastructure. That can make it easier to change providers or tools without rebuilding every workflow, while giving the company a way to monitor usage.
For engineers, the controls also shape how freely agents can be used. Spending tiers and manager approvals balance experimentation against costs, while human review remains part of the release process. Those arrangements matter as organizations assess whether AI assistance changes engineering throughput, review workloads and software risk. The source describes Uber’s systems and practices, but does not establish that they have improved productivity by a particular amount.
As an affiliate, we earn on qualifying purchases.
Infrastructure Built Before AI Agents
Uber’s internal tooling reflects a history of building for operational demands that commercial products did not meet at the time, according to the report. Medisetty joined in 2015, when Uber had two services, the article says; by 2019 it had thousands. The report says the service count in 2026 is expected to be roughly similar to 2019, despite the business being about ten times larger than it was then. Those figures are presented by the source and are not accompanied here by an independent methodology.
The company’s earlier systems include uOwn, which identifies service ownership; M3DB, developed to handle high volumes of metrics; Cadence, a durable workflow engine; and Morpheus, a platform for feature flags and experimentation. The article says related open-source work helped underpin Chronosphere and Temporal. It also describes cities as a core operational concept that affects infrastructure, including service monitoring, deployments and regional placement. This existing platform provides context for Uber’s newer effort to connect agents with code and internal engineering knowledge.
In 2023, Uber’s platform team decided that AI-built systems should be independent of any one model or harness, according to the interview. The Model Gateway and AIFX reflect that design choice. The report also discusses background agents and internal tools such as Slate, Code Inbox and DevPods, but the supplied source material does not fully specify the functions of every tool.
“All PRs are still reviewed by engineers.”
— The Pragmatic Engineer, describing the interview with Uday Kiran Medisetty
As an affiliate, we earn on qualifying purchases.
Productivity Results Remain Unspecified
The source describes Uber’s systems through an interview and does not provide independent validation of their effects. It is unclear how many engineers use each tool, how often agents produce code that reaches production, or whether the review tools have changed defect rates, delivery speed or reviewer workload. The reported roughly 50% token-cost reduction lacks a stated baseline and measurement window in the supplied material, so it should not be read as a general cost comparison.
The account also does not detail the rules behind risk scores, the scope of AI Guard’s checks, or how Uber evaluates the quality of agent-generated changes. The provided excerpt ends partway through its description of the code-review workflow, leaving some implementation details unavailable.
As an affiliate, we earn on qualifying purchases.
How Uber Measures Agent Use
The published report is an interview and does not announce a product launch, a rollout date or a formal next milestone. The next useful evidence would be fuller operational data from Uber: how it measures engineer productivity, what review outcomes look like, and how token spending and agent use change over time. For now, the account documents the infrastructure and controls Uber says it uses, while the results of those practices remain only partly quantified in the available material.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the news about Uber’s AI infrastructure?
The Pragmatic Engineer published an interview with Uber engineer Uday Kiran Medisetty about internal developer systems that connect AI coding agents to models, engineering tools and review workflows.
Which AI coding tools does Uber support?
The report says Uber supports Claude Code, Codex and OpenCode. Its AIFX wrapper is described as carrying settings, skills and plugins between supported harnesses.
Does AI replace human code review at Uber?
No. The report says engineers still review all pull requests. AI review and other information, such as risk scores and feature media when available, can help inform that process.
How does Uber manage AI costs?
According to the interview, engineers receive spending tiers, and managers can approve additional monthly spend when a tier is reached. A plugin called CostCheck offers cost suggestions across sessions.
Has Uber proved that these tools improve productivity?
The supplied report does not provide a measured productivity result. It describes the tools and practices, and reports a token-cost reduction of about 50% without giving the baseline or measurement window in the material provided.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
