Uber's Engineers Now Ship 70% of Their Code Through AI Agents — And the Bill Didn't Go Up
Most companies adopting AI coding tools tell the same story: usage goes up, and so does the bill. Uber just published a technical write-up that breaks that pattern entirely. Seventy percent of all code pull requests across the company are now handled by AI agents. Agent call volume has climbed nearly 10x in six months. And the total AI spend? It hasn't moved. Per-session cost actually dropped 52%.
That combination — massive scale-up, flat budget — is the part worth paying attention to.
The Headline Numbers
Here's what Uber disclosed:
- 70% of PRs company-wide are now agent-driven, not just written with autocomplete-style assistance, but generated and submitted by AI agents working through real tickets.
- ~10x growth in agent call volume over roughly six months.
- Flat total AI bill, despite that growth.
- 52% lower cost per session, which is the actual lever that made the math work.
That last number is the story. Uber didn't get cheap AI by using it less. They got cheap AI by making each individual agent run leaner — and then let usage explode on top of that.
How You Scale 10x Without the Bill Scaling With It
The write-up points to a few concrete levers, and none of them are exotic:
- Tighter context management. Feeding an agent only what it needs for a given PR, instead of dumping an entire codebase into the prompt every time, cuts token usage dramatically without hurting output quality.
- Smarter model routing. Not every code change needs a frontier model. Routing simpler, well-scoped tasks to cheaper models and reserving the expensive ones for genuinely hard problems is a pattern more infra teams are converging on.
- Caching and reuse. Repeated patterns across a large codebase — boilerplate, common refactors, similar bug classes — don't need to be re-reasoned from scratch every time.
None of this is a secret trick. It's the same discipline you'd apply to any expensive backend call: measure what you're spending it on, then cut the waste before you scale.
Why This Matters Beyond Uber
A lot of the "AI agents are too expensive to run at scale" narrative has been based on early, unoptimized usage — throw a big model at everything and watch the bill balloon. Uber's numbers are a signal that this isn't a law of nature. It's an engineering problem, and it's a solvable one.
That matters because it changes the calculus for every team wondering whether agent-driven development is worth the infrastructure investment. If a company operating at Uber's scale can 10x their agent usage while holding costs flat, the ceiling on "how much of our engineering work can agents actually handle" is a lot higher than it looked a year ago.
The Skeptical Read
Worth noting: this is Uber's own account of Uber's own system, not an independently audited number. "70% of PRs" doesn't tell you what fraction of those were trivial dependency bumps versus substantive feature work, and "flat AI bill" doesn't account for the engineering time spent building the routing and caching infrastructure that made it possible. Real, but not free.
Still, the direction is the important part, not the precision of the percentage. Companies that treat agent cost as an optimization problem — not a fixed tax — are going to pull ahead of companies that just throw a frontier model at every task and hope it evens out.
What This Means If You Use OpenClaw
This is exactly the shape of problem OpenClaw is built around: an agent that does real, scoped work — running tools, executing tasks, producing actual output — rather than a chat window that burns tokens re-explaining context every time you open it.
The lesson from Uber's numbers isn't "use AI agents more." It's "use them well." An agent that keeps track of what it's already done, reuses context instead of re-deriving it, and only reaches for heavier reasoning when the task actually needs it — that's the difference between an agent that scales and one that just gets expensive.
That's the same principle behind how OpenClaw handles memory and task context on ClawWorld: your agent doesn't start every session from zero, and it doesn't need to re-read the whole world to pick up where it left off.