Uber Lets Agents Validate Themselves: The Trick Is to Throw Them into a Real Staging Environment and Run Them Through It
How do you verify code written by an agent? Uber's approach is not to have the agent review its own code, but to deploy its branch into the SLATE isolated environment, run a comparison between main and the branch, and attach regression screenshots to the PR. This is done for Web, iOS, Android, and backend APIs.
Uber Lets Agents Validate Themselves: The Trick Is to Throw Them into a Real Staging Environment and Run Them Through It
Uber's Distinguished Engineer Uday Kiran Medisetty discussed their approach on The Pragmatic Engineer podcast. The gist of his words: Suppose I'm changing a feature that spans two or three services—how do I know it works end-to-end?
This problem existed before agents wrote code. Uber built something called SLATE to address it.
What Is SLATE
Simply put, engineers can create an instance for themselves that deploys the branch they are working on. Then you route the UID of a test account to this instance. When requests come in, they run your branch code, while upstream and downstream services still hit the original staging environment.
The effect: you can test your changes end-to-end without worrying about the rest of the infrastructure. This was the standard workflow for Uber engineers before agents appeared.
What Happens After Agents Are Integrated
The key point is here: Uber didn't build a separate validation system. They taught the agent how to deploy SLATE and the APIs of various services. Uber's reasoning: since this already exists, they just need the agent to learn how to deploy SLATE and give it the APIs to call these services.
Once the agent has these, it does four things:
- Runs an agent browser in the SLATE environment, visiting both main and the engineer's branch to perform A/B comparison
- Generates a complete before-and-after report, pointing out any regressions
- Attaches screenshots and videos from the run to the PR as evidence
- The same mechanism covers Web, iOS, Android, and backend APIs
Note the third point. This isn't the agent saying 'I tested it.' It's a replayable run record sitting in the PR.
Why This Design Is More Reliable Than 'Letting the Agent Self-Test'
One detail worth mentioning: when an agent writes its own tests, the tests and the changes often share the same blind spot—situations it didn't think of won't be covered by the assertions it writes.
Someone pointed out in a reply: the before-and-after report against main is the most valuable part, because it catches regressions the agent never thought to test for. Running a real branch in a real staging environment is what takes this beyond 'self-review.'
This is also what I think is worth noting. When most teams discuss agents writing code, the focus is on generation quality. Uber puts effort into the validation pipeline: environment isolation, traffic routing, comparison baselines, and evidence retention. When the generation-side model changes generations, the validation pipeline doesn't need to be rebuilt from scratch.
Additional Notes
SLATE is not a new thing built for agents. It originally solved a human problem: how to validate cross-service changes locally. The agent simply reuses it. This point is more worth remembering than the technical details: first have a reproducible validation environment, then talk about letting agents validate themselves.
There's more about Uber's AI Infra in The Pragmatic Engineer podcast:
https://x.com/GergelyOrosz/status/2100549218117022176
Original tweet:
发布时间: 2026-11-16 17:26