TL;DR
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
The Pragmatic Engineer reported on Uber’s developer tools and AI infrastructure in an interview with Distinguished Engineer Uday Kiran Medisetty. The account describes a shared model gateway, support for multiple coding-agent harnesses, spending controls and continued engineer review of pull requests; several scale and productivity figures are not fully specified.
The report says Uber’s Model Gateway, built in 2023, lets internal AI tools use models from Anthropic, OpenAI and other vendors, as well as open models Uber runs. The platform team’s stated aim was to keep AI development independent of a particular model or harness. The gateway also supports security checks, spending records, audit logs and session traces.
Uber’s AIFX, short for AI Foundations, wraps several coding-agent harnesses, including Claude Code, Codex and OpenCode. According to the report, it carries settings, skills and plugins between them. A separate security middleware called AI Guard checks requests and responses and tracks latency, while the CostCheck plugin offers engineers suggestions for managing usage.
Spending is managed through allocated tiers: engineers who reach their limit can seek a manager’s approval for more monthly usage. The report says Uber has cut token costs by about 50% in recent months, but the supplied material does not give the exact measurement period or baseline. Human engineers still review every pull request; internal tools provide AI review, feature screenshots or videos where applicable, and risk scores to help reviewers judge the work.
How Uber Is Governing Coding Agents
The systems described show how a large engineering organization is connecting AI assistants to existing software and internal knowledge while keeping controls around access, cost and review. A shared model gateway can give teams a way to change providers without rebuilding every integration, while centralized records and security checks give Uber visibility into AI use.
The approach also reflects a practical limit: faster code generation does not remove the work of evaluating changes. The report says code review has become a bottleneck at Uber, and that engineers remain responsible for reviewing pull requests. Tools that surface risk and test evidence may help direct reviewers’ attention, though the account provides no measured evidence that review times or defect rates have improved.
As an affiliate, we earn on qualifying purchases.
Infrastructure Built Before Generative AI
Medisetty joined Uber in 2015, when the company had two services, according to the report. By 2019 it had thousands. The article says Uber expects its service count in 2026 to be about the same as in 2019, while the business is roughly ten times larger than it was then. Those figures are presented in the source without a detailed definition of service count or the business-size measure.
The report connects Uber’s AI readiness to infrastructure built for earlier scaling problems. It cites uOwn for service ownership, M3DB for handling metrics, Cadence for durable workflows and Morpheus for feature flags and experimentation. It also describes city-level requirements embedded in systems such as observability and deployment: each city is associated with a cloud region and can move to another region. The interview presents these as examples of internal tooling developed where the company’s needs exceeded available products.
“Code review has become a big bottleneck inside the company.”
— Uday Kiran Medisetty, Uber Distinguished Engineer, as summarized by The Pragmatic Engineer
developer security middleware software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Metrics Still Need More Detail
The supplied report excerpt does not provide a publication date, the full interview transcript, or the complete set of its promised 11 learnings. It also does not specify the baseline and measurement window behind the reported 50% reduction in token costs, or how Uber calculates productivity as engineers use agents.
Details of the code-review risk scores, how often managers approve spending above an engineer’s tier, and the measured effect of these tools on review speed or software quality are not included. The account describes Uber’s internal systems but does not provide independent performance or security evaluations.
As an affiliate, we earn on qualifying purchases.
Further Evidence on Results
The source points readers to the episode transcript and timestamps on The Pragmatic Engineer’s report page, where the full interview may clarify implementation details omitted from the supplied excerpt. It does not announce a specific upcoming launch or milestone. Further reporting would be needed to establish how Uber measures agent-driven productivity, review bottlenecks and token savings over time.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did The Pragmatic Engineer report about Uber?
It described Uber’s internal developer infrastructure for AI agents, based on an interview with Distinguished Engineer Uday Kiran Medisetty. Topics included model routing, coding tools, spending controls and code review.
Which AI coding tools does Uber support?
The report says Uber’s AIFX wrapper supports Claude Code, Codex and OpenCode, and carries settings, skills and plugins between them.
Does AI replace human code review at Uber?
No. The report says engineers still review every pull request. AI review, screenshots or videos where relevant, and risk scores are used to support that process.
How does Uber manage AI spending?
Engineers receive a spending tier and can ask a manager to approve additional monthly usage after reaching its limit. The report also describes CostCheck, a plugin that offers usage-cost suggestions.
How much has Uber cut token costs?
The report says token costs fell by about 50% in recent months. The supplied material does not state the measurement window or comparison baseline, so the figure cannot be independently assessed from this excerpt.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
