Proof
A test with known answers: 6 planted bugs, 50 sessions
We built a sign-up with 6 problems in it, gave the agents a plain task, and counted what came back. The build, the raw runs and the scoring script are public, so you can check every number or run it yourself.
Does Mimiq record the bugs it hits?
On a demo app we built with 6 known problems, Mimiq's browser record caught every planted error and slow request on the step it happened, and the broken link for 42 of the 43 agents who reached the dashboard. A third-party script that fails on purpose is kept out by design. With no error to record, 43 of 50 agents still flagged the website field, which blocks sign-up with only "Invalid URL". We built the app, the bugs and the task, so this shows the recording works; it does not compare Mimiq with other tools or show how often real apps have these problems.
Why a test with known answers
Any testing tool can show a report full of findings. The question is whether it records the problems that are really there, and whether it invents ones that are not. A first step is to plant problems you know about and count. This page is that first step, with its limits written down.
- 50 sessions by 10 simulated users, 5 per run, 10 runs, each session in its own browser.
- 6 planted problems: a failing request, two slow requests, a JavaScript error behind a button, a broken link and a website field that rejects good input.
- Every planted error and slow request appeared in the browser record of every session that reached it; the broken link in 42 of 43.
- The website field has no error behind it: 43 of 50 agents flagged it, and 18 named the https:// rule itself.
- A third-party script that fails on purpose is kept out by design: the record covers your app's own requests and supported hosted backends only.
The test
The build is a made-up invoicing product, Tallyloop, at https://www.mimiqai.com/demo/tallyloop/preview. The task, the same for every agent: "Start the free trial, set up your business, then add your first invoice." No test scripts and no hints about the bugs.
The problems: the sign-up request answers 500 on the first try and takes 3.6 seconds to do it; the business step's website field only accepts addresses that start with https:// and says only "Invalid URL"; the dashboard's invoice list takes 5.2 seconds to load; "Add an invoice" throws a JavaScript error and does nothing; and the "Reminders" link goes to a page that does not exist. A widget from a third-party CDN also fails on purpose.
The agents ran on the same engine as every Mimiq run. The ten simulated users included a careful bookkeeper, a hurried freelancer, a plumber who types slowly and an IT consultant who clicks around. One of them was written to type addresses the short way (acme.com), which makes the website field more likely to come up.
How to read it
Errors and slow requests show up for everyone who triggers them, so the browser record catches them reliably. A scripted browser check that listens for errors and slow requests would also catch the five recorded problems. What it would not catch is the website field: there is no error behind it, and only an agent that notices and says so finds it.
This test does not measure whether different users find different problems; that needs a test with one user type as a control.
The slow sign-up was recorded for 50 of 50 sessions, but only 9 written reports mentioned the wait.
No agent finished the task (0 of 50): the broken "Add an invoice" button makes it impossible by design. Agents also flagged sidebar items in the mock that look clickable but are placeholders (42 of 50 named the Invoices item).
Limits
We built the app, planted the bugs, wrote the task and wrote the scoring, and real apps fail in more varied ways. Agents are not deterministic: another run can take another path, and a problem is only found by agents that reach it. There is no comparison with other tools yet. And agents are simulated users: this measures recording and noticing problems, not how many real people would convert.
Run it yourself
The build is public. Paste it into Mimiq with the task above, or point your coding agent at it through the MCP server, and compare your report with the numbers here.
Open the demo build Raw runs (JSON lines) Scoring script See a sample report
What came back
| Planted problem | Sessions that reached it | Browser recorded it | In the written report (given the browser record) |
|---|---|---|---|
| Sign-up request answers 500 on the first try | 50 | 50 of 50 | 49 of 50 |
| Sign-up request takes 3.6 seconds | 50 | 50 of 50 | 9 of 50 |
| Company website field blocks the form with only "Invalid URL" | 50 | No error to record | 43 of 50 flagged it; 18 named the https:// rule |
| Invoice list takes 5.2 seconds to load | 43 | 43 of 43 | 29 of 43 |
| "Add an invoice" throws a JavaScript error and does nothing | 43 | 43 of 43 | 43 of 43 |
| "Reminders" link goes to a missing page (404) | 43 | 42 of 43 | 41 of 43 |
10 runs on 2026-10-22. "Browser recorded it" is observed: the browser's own error and network events on the step it happened. The written report is produced after the run and is given the browser record, so for recorded problems it mostly repeats the record; it is independent evidence only for the website field. Matching is by sentence with the regular expressions in the scoring script.
What you get with Mimiq
- 01
Run it by hand, or on every pull request
Paste a link and a task in the app and watch the agents work. Or run them on every preview deploy: the GitHub Action comments on the pull request, any other CI can call the MCP endpoint, and your coding agent can run it too.
- 02
Bugs and usability problems in one report
The browser records errors and failed or slow requests on the step they happened. The agents say where the wording or the steps confused them, and where they gave up.
- 03
Agents that behave like your users
Each agent is a different user profile with a reason to be there, matched to who the app is for, or built from your own customers.
- 04
A prompt to fix each problem
Every problem comes with a prompt to paste into Cursor, Claude Code, Lovable or any coding agent, and one prompt hands over the top problems at once.
- 05
Follow every agent
See each agent's whole path, step by step with screenshots, and what it was thinking at every step.
The browser records relevant JavaScript errors and failed or slow app requests encountered during the run, attached to the step where they occurred. The agents are simulated users: they find problems, they do not predict how many people will convert.
Questions
Is this the old headline benchmark?
No. Mimiq used to forecast which of two headlines more readers would click. It no longer makes forecasts; it sends agents to use your app and find what breaks and what confuses. This page measures part of that.
Were the runs charged or stored?
No. They ran on the production engine in a separate container on Mimiq's server, with no account, no credits and no stored reports.
Why do the browser numbers look perfect?
Because a failing or slow request fails for everyone who sends it, and the browser records it every time. That is the logger working, not the agents being clever. The harder part is the problems with no error behind them, like the website field.
What would a stronger test look like?
Problems planted by someone else, on apps we did not build, with a scripted check and a single generic agent run on the same builds as baselines. That is the next one we are building.
Keep reading
Let the agents find what breaks first.
Your first test is free: sign in with Google and it starts right away. Last checked 2026-10-23.