TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A Pragmatic Engineer interview with Uber Distinguished Engineer Uday Kiran Medisetty details how the company’s existing developer systems and newer AI tools support engineers using coding agents. The discussion covers model access, security checks, spending controls and code review; it does not provide independent evidence that the tools have raised productivity.
Uber’s internal AI infrastructure includes a model gateway, tools for switching coding-agent harnesses and systems for tracking usage and spending, according to Distinguished Engineer Uday Kiran Medisetty in an interview published by The Pragmatic Engineer. The account describes how Uber is connecting AI agents to its code and engineering workflows while retaining human review of every pull request.
Medisetty, described by The Pragmatic Engineer as one of Uber’s longest-tenured engineers and a leader in its agentic AI work, walked through tools used by engineers, including Model Gateway and AIFX. Uber built Model Gateway in 2023 to keep AI-powered systems independent of any single model or harness, the report says. It supports models from vendors including Anthropic and OpenAI, as well as open models run by Uber.
The gateway also provides a point to add security checks and usage records. The report says the security team’s AI Guard middleware checks requests and responses, while the platform tracks spending by user and keeps audit logs and session traces. AIFX, short for AI Foundations, lets engineers use Claude Code, Codex or OpenCode while carrying settings, skills and plugins between harnesses.
Uber assigns engineers AI spending tiers; a manager can approve additional spending after an engineer reaches a limit. The report also describes CostCheck, a plugin that provides suggestions across sessions. Medisetty said Uber’s token costs had fallen by about 50% in recent months. The source does not specify the measurement period, baseline or calculation behind that figure.
How Uber Governs AI Coding
The tools show how a large engineering organization is trying to manage coding agents as part of its existing development systems, rather than treating them as standalone assistants. A shared model gateway can give platform and security teams a place to apply controls and record usage, while harness flexibility can help engineers change tools without rebuilding their individual setups.
Code review remains a human responsibility at Uber, according to the report, but the company is developing systems to help reviewers prioritize their time. Medisetty described code review as a bottleneck. AI review, risk scores and supporting materials such as screenshots or video are used to give engineers more information before they assess a change. The interview describes Uber’s approach, not a measured finding that these systems have reduced review times or improved software quality.
As an affiliate, we earn on qualifying purchases.
Infrastructure Built Before AI Agents
The interview places Uber’s agent tools against a history of building internal infrastructure for the company’s scale and operational needs. Medisetty joined Uber in 2015, when the company had two services, according to the report. By 2019 it had thousands; the report says the number in 2026 is expected to be roughly similar to 2019 even though the business is about ten times larger. These figures are presented in the source interview, with no independent verification or detailed definition of “services.”
Uber developed systems including uOwn for identifying service ownership, M3DB for metrics, Cadence for durable workflows and Morpheus for feature flags and experimentation. The report says these were created to address needs that available products could not meet at the time. It also describes the company’s city-based operating model: cities affect infrastructure decisions such as observability, deployments, service sharding and cloud-region placement.
That established platform is relevant to the AI discussion because agents need controlled access to code, services and internal engineering knowledge. The interview’s central point is that Uber’s AI tooling builds on infrastructure developed before generative AI became part of routine software work.
““Everything built with AI needs to be model- and harness-agnostic.””
— Uday Kiran Medisetty, as quoted in The Pragmatic Engineer’s interview
developer security and audit logs software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Interview Does Not Measure
The report is an interview and company-insider account, not an independently audited evaluation of Uber’s AI systems. It does not quantify how much agents have changed engineering output, defect rates, review times or delivery speed. The stated 50% reduction in token costs lacks a specified baseline and time window, so it cannot be compared reliably with other periods or organizations.
The source material also gives limited detail on how AI Guard evaluates requests, how risk scores are calculated, or what circumstances lead managers to approve additional spending. It describes the tools and policies but does not provide adoption figures, spending totals or results broken down by team or task.
As an affiliate, we earn on qualifying purchases.
Further Details on Agent Use
The Pragmatic Engineer says the episode examines how Uber engineers use background agents, how tools such as Slate support development and bug fixing, and how the company is adapting code review and productivity measures. The source page points readers to the episode transcript and recordings on YouTube, Apple and Spotify; it does not identify a further publication date or announced company milestone.
More evidence about the approach’s effects would require measurable results, including clearly defined cost comparisons, review performance and software quality. The interview does not say when Uber will publish such data, or whether the described tools and policies will change.
code collaboration and review software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Uber’s AI infrastructure?
As described in The Pragmatic Engineer interview, it includes Model Gateway for access to different AI models, AIFX for switching coding-agent harnesses, and tools for security checks, usage records and spending controls.
Which coding-agent harnesses does Uber support?
The report says AIFX supports Claude Code, Codex and OpenCode, carrying settings, skills and plugins when an engineer switches between them.
Does AI now review Uber’s pull requests instead of engineers?
No. The report says engineers still review all pull requests. AI review and other information, including risk scores and feature media where applicable, are intended to assist reviewers.
What does Uber’s reported 50% token-cost reduction mean?
Medisetty said token costs had declined by about 50% in recent months. The source does not specify the comparison baseline, exact time window or calculation, so the figure’s basis is unclear.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
