Browser-Based Agentic Testing
How an in-app browser turns a scattered testing session into evidence the next person can act on.
Product and platform · · 7 min read
Manual testing rarely fails because someone was careless. It fails because the session happened in a separate window, the state that produced the bug was never captured, and by the time it is written up the interesting details are gone.
A test is a claim about a flow
Checking a page is not really about the page. It is about a user in a particular state moving through an intended journey, and whether the result matches what the change was supposed to do.
Framed that way, a test has the same shape as any other product artifact: a claim, the evidence for it, and a conclusion. Sessions that do not produce those three things tend not to survive contact with the next person.
Authenticated state is where the product actually lives
Most meaningful behaviour sits behind a login: the populated dashboard, the permission-dependent view, the half-finished checkout. Testing that only covers logged-out pages covers the least interesting part of the product.
Persistent per-workspace browser state matters here more than it sounds. If every session starts by logging in again, people test the fast path and skip the deep one — which is exactly backwards.
If every test session starts with logging in again, people will test the shallow paths.
Capture the state, not the impression
“The button doesn't work” is a description of a feeling. What the next person needs is the URL, the account state, the console output, the network response, and what was expected instead.
The value of doing this in the workspace is that the capture is automatic rather than a discipline someone has to remember at the end of a frustrating session — which is precisely when nobody remembers.
What turns an observation into a usable report
- The exact state: URL, account, feature flags, viewport.
- What was expected, stated before the result.
- What happened, with console and network evidence attached.
- Where the next person should look first.
What agents are good and bad at here
Agents are reliable at the repetitive part: walking a known flow, checking it across viewports, capturing artifacts consistently, and doing it again after every change without getting bored.
They are weaker at noticing that something is subtly wrong in a way nobody specified — the layout that is technically correct and visually off, the copy that is accurate and confusing. That judgment is still human, and a testing workflow that pretends otherwise will miss the defects that matter most.
Key takeaways
- A test is a claim, its evidence, and a conclusion.
- Authenticated state is where the important behaviour lives.
- Capture state automatically; nobody documents well when frustrated.
- Agents handle repetition; humans still catch subtle wrongness.