Over 70% of Uber's Merge Requests Completed by Agents
Uber's engineering blog states that over 70% of code merge requests are attributed to local or cloud agents. Engineers have written over 3,600 agent skills covering the development lifecycle, executing more than 30,000 times daily.
From February to August 2026, the weekly active users of agent tools increased sevenfold, and the weekly request volume increased 9.4 times. The total AI billing has stabilized since April. After fixing the same model to strip its own optimization, the cost per thousand model requests dropped about 34% from the peak, and the cost per session decreased by 52% from the June peak.
More and more sessions are no longer initiated by humans. Managed agents handle code reviews, self-healing continuous integration failures, end-to-end merge requests with visual validation, duty alert triage, defect debugging, and daily maintenance, with humans only performing reviews and upgrades.
At the core is a unified gateway: all company models call through the same entry point, performing identity checks, desensitization of over twenty types of personal information, and security modeling, with daily requests exceeding 100 million and over 800 projects. The tool gateway projects internal interfaces to the command line, avoiding stuffing entire responses into context, which previously led to over a 40% reduction in fleet-level token consumption.
The skill market manages reusable processes, and the context graph replaces agents crawling through systems with tens of millions of entries. The code review engine uReview issues about 25,000 comments weekly, with approximately 67% being processed, reducing costs by about 60% compared to the initial rough implementation. The average number of lines of code produced per person has roughly doubled year-on-year, with automated migration reaching about nine million lines.
In terms of market mechanisms, the cost is borne by engineering budgets treating agent usage as factory capacity, while model calls are sold on a token basis. Events are driven by internal platformization, with funds flowing from "adding one more model" to routing, caching, skill reuse, and context compression; benefiting are platform teams that can design down session costs, while pressured are access parties that can only connect chat windows to warehouses but cannot control bills.
Source: Public Information
ABAB AI Insight
70% of merge requests are not trophies from prompt competitions, but indicators of factory utilization. Uber has written agents as measurable production lines: skills are processes, gateways are quality checks, and session costs are unit product costs. The 9.4-fold increase in requests while billing remains flat indicates that scarcity has shifted from model interfaces to cost equations. Those who cannot solve this equation will blow their budgets once usage rises.
The capital path is for engineering organizations to shift model costs from departmental reimbursements to platform depreciation. Money is no longer distributed based on "which large model this group used," but rather pressed based on every thousand requests, each session, and each review. Routing to small models, caching to align idle intervals, and tool modes to batch cut tokens all transform agents from consultants into piece-rate workers. Humans shift from writing code to writing skills and upgrade rules.
The benchmark is automotive factories writing robot density into unit hours, and cloud vendors writing virtual machine utilization into gross profit. Software is transitioning from "programmers plus assistants" to "agents plus reviews": capacity constraints are shifting from staffing to prioritization. The ability to do something is no longer the bottleneck; the question is whether it should be done.
Structurally, this belongs to technological substitution. The mechanism is: when merge requests can be produced in bulk by agents, human hours shift from input to governance. Accessing models only solves usability; designing costs solves scalability. The definition of a factory is that usage can run rampant, but unit costs must be pre-written into the system.
ABAB News · Cognitive Laws
Usage can run rampant, provided costs are designed first.
Accessing models is just starting work; managing sessions is the factory.
Humans shift from writing code to writing skills and upgrade rules.