Agentic testing
what is agentic testing?
Agentic testing means an autonomous agent decides how to exercise an application, rather than replaying steps a person wrote down. You describe the outcome; the agent reads the page, chooses actions and judges the result. It removes selector maintenance and adds two new questions: what a run costs, and whether the agent's judgement of success can be trusted.
The appeal is maintenance. Traditional automation encodes how to do something, and how changes constantly, so the suite decays. Encoding what should happen instead means a renamed button, a moved form or a redesigned page do not invalidate the test.
The first honest cost is money and time: reasoning about a page on every run is slow and billed per token. Most serious implementations answer this by recording a successful run and replaying it without a model, which makes the ordinary case cheap and reserves the agent for when something changed.
The second is trust, and it is the harder one. An agent that decides whether a run passed is grading its own work, so the useful implementations tie the verdict to something checkable: a named element, quoted page text, an artefact a human can open. A green tick with no evidence behind it is a worse signal than a red one you can read.
in smolanalytics
The agent reads the page through its accessibility tree, picks named controls, and returns one of five verdicts. A failure quotes what it saw instead of what the sentence asked for, names the file most likely responsible and the evidence connecting it, or says plainly that it has no candidate rather than guessing. Passing runs are recorded and replayed with no model call, so the agent is the exception rather than every run.
How it works shows where this fits in the loop, and the docs have the exact behaviour.