The AI agent in this product does not chat for the sake of chatting: it does concrete work — analyzes comments, writes and fixes actions, repairs broken checks. The chat is powered by a separate program, Xedant Agent, and everything the agent does is visible and can be reread. Below is what it can do and what happens when it is not connected.
How to connect it
The chat works through an external Xedant Agent: at installation you give its address and key. The test-manager skill ships with the product and syncs itself into the agent at startup — it is what explains to the model how to write checks and what to answer to a screen report. Nothing needs extra configuration.
One more thing worth knowing: the same agent that analyzes comments can also edit the project’s files. That is why its address and key are as valuable as a password, and should be shared only for good reason.
What the agent does
- Analyzes comments — the main job. From your phrase “the button is invisible on the dark background” comes a check that from then on looks for this problem by itself.
- Writes actions — from a description in words: “sign in and open the payment page”.
- Repairs checks — at your request, and on its own when a rule or action broke during a run.
- Helps with visual differences — works through the ambiguous cases, where the simple comparison rules gave no answer, and explains the difference in words.
- Sorts failed builds — gathers identical failures into groups by cause and explains what the snapshots in a group have in common.
- Proposes covering tests — looks at the list of what is not checked yet and suggests what to close the gap with.
- Proposes a test healing — when a test broke, prepares a new revision of the file.
In every case the result is not text in a chat but a file in the project: a rule, an action, a test or an edit. And one rule covers all of it: AI proposes, a human decides. Classifying a difference, a verdict on a group of failures, a new test, a healing — all of these are proposals only; the last word always stays with a person.
What it looks like
- The chat panel on the right — opens next to the project without getting in the way of the lists.
- Model and skill selection — a fast model is enough for simple analyses; for complex ones a stronger model can be picked. The
test-managerskill is chosen for comment analysis. - The analysis history — with search: a month later you can find what exactly the agent proposed and why.
- Queues and labels — when there are many analyses, they queue up; labels help sort them by topic.
Budgets and the model
A project can say which model its visual tasks run on and set spending ceilings: how many tokens one task gets, how many a whole campaign gets, how many healing attempts are allowed. These are circuit breakers: they stop the spending as it approaches the ceiling but never approve anything. Approving a difference, as before, stays with a human.
What was spent and on what is visible in the AI log; the ceilings themselves are edited in the Visual Testing Settings section.
A ready report instead of explanations
You do not need to explain the situation to the agent. The “Call agent” button on a screen places a full report into the chat: the comments verbatim, a link to the snapshot, the knowledge accumulated about the project, the list of rules and the screen types. All that is left is to add your request and send it. This is the fastest path from “I do not like this button” to a working check.
What happens without the agent
When the external agent is not connected, the chat is simply hidden — and that is honest, because it could not work. Comment analysis, action generation and auto-fix are unavailable along with it. Everything else works in full: screens, rules, actions, tests, runs, schedules, issues and the summary. Rules that are already written keep checking pages — no AI is involved in them at all.
And about the license: it is needed for exactly one thing — sending a message to the agent. The product itself, including all checks and runs, works without it. The details are in the Licensing section.
Next: how to show the agent a real open page — in the Live Page & MCP section.