AI Writes Your Code Now. Incidents per Pull Request Are Up 242%.
AI-generated code is outrunning code review: incidents per PR up 242%, bugs per developer up 54%.Why end-to-end tests are now the quality gate that scales.
10 min readBy the GUI Test Studio team
What Faros AI measured
In April 2026, Faros AI published The AI Engineering Report 2026: The Acceleration Whiplash.It is not a survey of how developers feel.It is about two years of telemetry from 22,000 developers in more than 4,000 teams, and it compares the periods of lowest and highest AI adoption inside each organization.
The first finding sets the scene.In 80% of the teams, more than half of the developers use AI tools every week, and the share of AI-generated code that developers accept rose from 20% to 60%.
“AI is not assisting developers. In most organizations, it is leading them.”
The good news is real, and the report says so.With high AI adoption, teams complete more work.
+66%
Epics completed per developer
+33.7%
Tasks completed per developer
+16.2%
Pull requests merged per developer
Then comes the whiplash.The same teams ship more incidents and more bugs, and the curve is getting steeper, not flatter.
Production quality with high AI adoption
Incidents per pull request
Incidents per merged change occur at more than three times the rate.
Monthly incidents
Outages, security events and failures that reach users.
Bugs per developer, 2026 report
Bugs per developer, 2025 report
The defect rate is steepening, not flattening.
Bugs per pull request
Why are bugs and incidents rising with AI-generated code?
An incident is what a user feels: an outage, a security event, a system failure.Between low and high AI adoption, the ratio of incidents to merged pull requests rose 242.7%, and monthly incidents rose 57.9%.
“For every PR merged, incidents are occurring at more than three times the rate.”
The bugs follow the same line.Faros's 2025 report, The AI Productivity Paradox, measured bugs per developer up 9%.In this dataset, the figure is up 54%, and the report's own page puts bugs per pull request up 28.7%.The more a team relies on AI, the faster the defects grow.
These numbers describe what happened, not one single cause, and Faros is careful to say so.But they all point the same way.More code reaches production, faster, and each change carries more risk than before.
Why can't code review keep up with AI coding agents?
Code review was designed for code written at human speed.An AI coding agent writes a pull request in minutes, and a team with five agents opens more pull requests in a day than its senior engineers can read in a week.
Code review under AI-generated volume
Median time in review
Median time to a first review
Pull requests merged with no review at all
Neither a person nor an AI agent looked at them.
Faros measured the queue: the median time to a first review is up 156.6%, the median time in review up 441.5%.And the pressure has an outlet.Pull requests merged with no review at all, human or agentic, are up 31.3%.
“Code is reaching production systems with no oversight at a meaningfully higher rate.”
Even a careful review has a blind spot.
- A reviewer reads a diff, and a diff never shows the running application.
- A button hidden by a new style, a total computed the wrong way on the checkout page, a dialog that opens behind the window: none of them is visible in the code that caused it.
What catches the bugs that review misses?
The first idea is to ask the agent for more unit tests.Do it, but do not stop there.Tests written by the agent that wrote the code share its understanding of the task: if it misread the requirement, the code and its tests agree with each other, and both are wrong.
The safety net that matches AI's pace has three properties.
| Property | Why it matters with AI-generated code |
|---|---|
| Automated | It runs on every pull request and every build, at the agents' speed, without waiting for a reviewer. |
| Independent | A person owns it, and it does not depend on the code or on the AI that wrote the code, so it cannot share their mistakes. |
| Checks what users see | It drives the real application through its interface and verifies the screen, where incidents are felt. |
That is an end-to-end test: a test of the whole application through its interface, the way a user goes through it, from sign-in to payment.It is what testing AI-generated code comes down to, and it is what the Faros numbers ask for: a check that grows with the volume of changes, not with the number of reviewers.
An agent opens a PR
Minutes of work, many files
Unit tests pass
The agent's own understanding
End-to-end tests run
What the user sees, owned by a person
Merge, or blocked
A failure waits for a fix
How do you make end-to-end tests a CI/CD quality gate?
Definition
A quality gate is a check a change must pass before it is merged or deployed.
- Five habits turn end-to-end tests into one an agent cannot talk its way around.
- 1
List the journeys that must never break
Sign-up, sign-in, search, checkout, the report your customers export every Monday.Five to ten journeys, each with the outcome a user expects.That list is the specification, and it belongs to a person, not to the agent.
- 2
Test the screen, not the code
Verify what a user reads and sees: the total, the message, the button, the arrangement.An agent renames classes and rewrites components freely.A test anchored on the screen survives that, and still fails when the screen is wrong.
- 3
Run on every pull request, and block on failure
A result that only warns is soon ignored.Make a failed end-to-end test fail the build, so the change waits for a fix instead of reaching users.
- 4
Keep the tests out of the agent's reach
The agent may run the tests and read the results.A change to what a test expects needs a person's approval, or the tests drift toward whatever the code happens to do.
- 5
Make failures readable, and file them
A failure should say which step failed and how the screen differs, with pictures, so the agent can propose a fix and a person can decide.And it should become a bug in your tracker without anyone copying screenshots: see turning failed tests into Jira bugs.
One warning: a gate that fails for no reason is worse than no gate, because people learn to ignore it.A gate needs tests that tell a real regression from harmless noise.That is the subject of flaky tests and false positives.
Do
- End-to-end tests of the journeys users rely on
- A failed run that blocks the merge
- A person who approves every change to what a test expects
- Failures with pictures and steps, filed as bugs
Avoid
- The agent's own tests as the only safety net
- Merging when the reviewers are too busy to look
- A red end-to-end run that only warns
- Gates so flaky that nobody trusts them
Testing AI-generated code on every build with GUI Test Studio
GUI Test Studio is built for exactly this gate.It tests web apps, Windows desktop apps, phones and games end to end, through their interface, with no test code to write or maintain.
A test in minutes: record it, or describe it
Record a workflow once in GUI Test Runner-Recorder, clicking and typing through the real application.Or describe it in a prompt to Claude, ChatGPT with Codex, or a local AI model, one we provide or your own, and the test is created for you.Every check an assistant adds is tried on the spot and refused if it could not pass on the screen right now.
Either way, the test checks what the screen shows.
- A Visual Assert compares the picture of an element, a Text Assert reads its words.
- The GTS Engine + AI finds each element by asking the application first, through its accessibility information or the web page's own elements, and by sight when the application does not say.
- A renamed class or a rewritten component does not break the test.
A gate in your pipeline, with gtr-cli
gtr-cli runs the same tests on a Windows build machine and answers the way a pipeline expects: its exit code fails the build, a JUnit file fills the test tab, and a Markdown summary is ready for the pull request.With one option, a failed test becomes a Jira issue with the steps to reproduce and the expected and found pictures.
gtr-cli run tests --countdown 0 --format github --junit results\junit.xml ^
--summary results\summary.md --artifacts results\artifacts --jira on-failureIt works in TeamCity, Jenkins, GitHub Actions, GitLab and Azure DevOps: see Run tests in your pipeline, and Run your end-to-end tests on every build for a TeamCity setup step by step.
A real regression, or harmless noise?
Every run ends in a Confidence Report: for each check, whether the element was found in the Search Phase and whether it matches what was recorded in the Verification Phase.
- When a picture differs, AI Smart Verification takes a second look.
- It ignores rendering noise, and still catches a changed word, a missing badge or a button that turned gray, then says what changed in plain words.
The agent can run the gate too
Through its connector, the assistant that writes your code can run your tests, read the verdict and propose a change to a test, which waits for your confirmation.A person stays in charge of what the tests expect.The whole cycle is in Development with AI agents (ADLC).
Key takeaways
- 1AI is the primary author of code now. Throughput is up, and so are incidents: +242.7% per merged pull request, +57.9% per month.
- 2Bugs are accelerating, not stabilizing. Bugs per developer rose 54%, against 9% in the 2025 report.
- 3Review has hit its limit. 31.3% more pull requests are merged with no review at all.
- 4The safety net must be automated and independent: end-to-end tests of what users see, owned by a person, run on every build.
- 5Make it a gate. A failure blocks the merge, files a bug, and says what changed.
Questions
Does AI-generated code cause more bugs?
Faros AI's AI Engineering Report 2026 compared the periods of lowest and highest AI adoption inside each organization, across 22,000 developers.With high adoption, bugs per developer rose 54% and incidents per merged pull request rose 242.7%.The report measures what happened, not a single cause, but the defect rate rises as adoption deepens.
Why is code review not enough for AI-generated code?
Review time grows with the volume of code, and AI multiplies the volume.Faros measured the median time to a first review up 156.6% and pull requests merged with no review at all up 31.3%.A reviewer also reads a diff, never the running application, so a broken screen can pass a careful review.
How do you test AI-generated code before it reaches production?
Keep end-to-end tests of your critical workflows that a person owns, independent of the code and of the agent that wrote it.Run them on every pull request or build, and let a failure block the merge.They check what users see, whoever wrote the code.
What is a CI/CD quality gate?
A quality gate is a check in your pipeline that a change must pass before it is merged or deployed.A failed end-to-end test fails the build, so the change waits for a fix instead of reaching users.
Can an AI coding agent create and run GUI Test Studio tests?
Yes.Claude, ChatGPT with Codex, Copilot Chat in VS Code, or a local AI model, one we provide or your own, can create a test from a prompt and run it through GUI Test Studio's connector.In a pipeline, gtr-cli runs the same tests without any assistant.
Figures: Faros AI, The AI Engineering Report 2026: The Acceleration Whiplash (April 2026), and its report page for bugs per pull request.Faros AI, Claude, ChatGPT, Codex, Copilot, OpenAI, Anthropic, Jira, TeamCity, Jenkins, GitHub, GitLab, Azure DevOps and YouTube are trademarks of their respective owners.
Ready to record your first test?
Every feature, for 15 days.
Rather see it on the kind of app you test?Book a free 30-minute call with our team.