What stops the agent from inventing buttons?

Before any plan is generated, the page scanner fetches the URL and extracts the real navigation, buttons, inputs, and headings. Steps that target elements not present on the page are flagged, stripped, or rejected before they can bill a run.

Does Well Tested run without approval?

No. Nothing runs until you approve the plan. Once you keep a test, it reruns those approved steps on every push, every deploy, a schedule, or when your coding agent asks. A step repaired inside a run is labelled, and the repair is only kept if you approve it.

Can the agent modify production code?

No. The agent runs tests; it does not commit code. When a run fails you get a fix prompt to hand to your own coding agent.

What does the evidence contain?

Each run returns screenshots per step, a video of a failure, browser console and network logs, a diff against the last green run, and a shareable proof page. Failed runs also get Jev + Clef triage (real bug or test/environment issue) and a fix prompt. Any test can be exported as standalone Playwright code.

Does it work with CI/CD?

Yes. Kept tests rerun on every GitHub push, every deploy (deploy hook URL) or a schedule, and a CI endpoint returns a stable exit code and an evidence link.

Definition

What is agentic testing?

Tell Well Tested what should work. It scans the target, drafts a test for approval, runs it in real browsers, and returns replayable evidence. The agent does the work; you keep the authority.

The definition

Agentic testing is software QA in which an AI agent does the work a tester used to do by hand: it reads the live page, proposes the steps of a test, runs them in a real browser and reports what passed and what did not. The agent owns execution. You own the intent and the approval.

That split is the whole idea. The AI proposes, a human approves, and every run returns evidence a person can check. Nothing the agent drafts runs until you have read it.

How agentic testing works

Scan. Before any plan is generated, the page scanner fetches the target URL and extracts its real navigation, buttons and form fields. Hallucinated targets are blocked at this gate.
Plan. The planner only references elements that exist on the page: visible text, role and name pairs, or stable URL paths. You get a proposed plan you can read and edit.
Approve. Approving locks that exact version of the plan. First runs are supervised; only tests you approved and kept rerun on their own, on every push, deploy or schedule you choose.
Evidence. Each step captures a screenshot, the assertion outcome, and the browser console and network logs. Failures get Jev + Clef triage and a fix prompt. The share URL outlives the run.

Agentic testing vs test automation

Test automation needs a human to record or script every step. Agentic testing drafts the test from intent and a live page scan. Side by side, the difference is who does the work.

Scripted test automation compared with agentic testing
	Scripted test automation	Agentic testing
Who writes the steps	A human writes or records every step	You describe the intent; the agent drafts the plan from a live page scan
When the UI changes	Selectors go stale and the test breaks	The step heals inside the run and is labelled "Passed after repair"; you decide whether to keep the repair
Coverage	Plateaus as the suite grows	Tracks the customer journey, not the test list
Maintenance	The biggest line item	Repair proposals arrive as versions you approve
Lock-in	Tied to the framework you scripted in	Any test exports as standalone Playwright code

Record-and-playback tools capture literal selectors. When a label changes or a button moves, an agentic step heals inside the run; if no locator matches, a vision fallback clicks what it can see and is labelled “Clicked by vision”.

It does not have to replace your existing automation. Use it where intent-driven coverage matters most, such as sign-up flows, marketing pages and customer journeys, and keep deterministic API suites where they earn their keep.

For apps built with Cursor, Claude Code and Lovable

Vibe coding is shipping features by directing an AI coding assistant in plain language. The code changes by the hour; scripted tests cannot follow. Agentic testing proposes steps from the live page, and kept tests heal small UI changes inside the run, so coverage keeps up with the journey instead of the test list.

Your coding agent can run the tests itself. Run npx welltested setup once and Claude Code, Cursor, Codex, Windsurf, Copilot or Kiro get the Well Tested MCP server and a verify-before-done skill. Before it says done, the agent reruns your kept tests in a real browser; a failure comes back with the Jev + Clef reason and a fix prompt. Passing runs use credits; a run that catches a bug is free.

When typing breaks your flow, speak the goal instead: up to 60 seconds, English or Turkish. The transcript fills an editable goal you review before the first run.

Frequently asked

What is agentic testing?: Agentic testing is software QA in which an AI agent scans the live target, drafts the test, executes it in real browsers, and reports replayable evidence, under explicit human approval.
What stops the agent from inventing buttons?: Before any plan is generated, the page scanner fetches the URL and extracts the real navigation, buttons, inputs, and headings. Steps that target elements not present on the page are flagged, stripped, or rejected before they can bill a run.
Does Well Tested run without approval?: No. Nothing runs until you approve the plan. Once you keep a test, it reruns those approved steps on every push, every deploy, a schedule, or when your coding agent asks. A step repaired inside a run is labelled, and the repair is only kept if you approve it.
Can the agent modify production code?: No. The agent runs tests; it does not commit code. When a run fails you get a fix prompt to hand to your own coding agent.
What does the evidence contain?: Each run returns screenshots per step, a video of a failure, browser console and network logs, a diff against the last green run, and a shareable proof page. Failed runs also get Jev + Clef triage (real bug or test/environment issue) and a fix prompt. Any test can be exported as standalone Playwright code.
Does it work with CI/CD?: Yes. Kept tests rerun on every GitHub push, every deploy (deploy hook URL) or a schedule, and a CI endpoint returns a stable exit code and an evidence link.