Qarmiq combines human testing, autonomous AI agents, and repeatable automated checks to verify critical product behavior. Every finding builds on what your team already knows about the product.
From a customer signing in to an order generating an invoice, every product has critical behaviors that must keep working. Qarmiq records these expectations in plain language and turns them into checks that can be verified as the product changes.
Qarmiq captures the result and supporting evidence.
A separate AI judge evaluates the recorded evidence and determines whether the check passed, failed, or remains inconclusive.
Behaviors that have been verified repeatedly can become Playwright scripts, giving subsequent runs a repeatable starting point.
Each lane works alone. They compound because all three record against one scenario.
Every finding connects the failure to the behavior that should have worked. Qarmiq captures evidence and reproduction steps so developers and coding agents can investigate the issue with context.
No agent setup and no engineering lift. Witness is one script tag in your app. Testers test as they normally do and mark pass or fail, and the verdict is written straight into the run. Defects are filed in place with an annotated screenshot, and the evidence (clicks, network calls, console errors, performance) is recorded while they work. Typed values are never recorded.
What a scenario looks like →AI coding tools that support MCP connect to Qarmiq’s MCP server with an API key. They read project and defect context, write scenarios and checks, ask for suggested scenarios, file defects and create runs. Every write is planned, confirmed and logged, and nothing can be deleted through it. A defect arrives with the evidence and a developer brief, so the agent starts from what happened.
How findings keep a suite honest →Choose Test autonomously on one check or a whole run and an agent tests it in a real browser. A separate judge rules pass, fail or inconclusive from the evidence, and an answer that cites none counts as inconclusive. A pass becomes a Playwright script. What the agent learns is proposed on the Decision Board, and waits for a person. Today a person starts each run; starting one from CI is on the roadmap.
Why a verdict can be trusted →Each level is defined by who leads, not by what kind of testing happens. Most teams begin at 01, connect their coding agent when developers ask, and delegate further as the workload grows.
Only the trigger and the inputs differ.
| Job | Tester in the app | Coding agent over MCP | Autonomous agent |
|---|---|---|---|
| Passing or failing a check | Witnesspass or fail, per browser | Creates the runand reads the verdicts back | Agent and judgecited evidence or inconclusive |
| A defect, with evidence | Filed in placeannotated screenshot, test log | File a defectand read the developer brief | Trace and videotyped text masked |
| New scenarios and checks | Written in plain sentencesno clicks, no selectors | Author and add checksplanned, confirmed, logged | Proposedaccepted by a person |
| Automation | The check stays with a person | Reads how a scenario is testedthe scenario playbook | Verified Playwright scriptdownloadable |
| Knowing if it is stable | Their verdicts count in it | Reads the project context | Their verdicts count tooone reading, set by rule |
Each one can be explained in a sentence, and all of them share the same record.
One plain sentence about what must always be true, and what proves it
Marks a scenario non-negotiable. A failing one opens an alert
A fixed set of scenarios, one environment, a verdict per check and browser
A widget in your own app that records verdicts and evidence as testers test
An agent tests in a real browser and a separate judge rules
Playwright scripts proven by repeated runs, updating themselves
Everything Qarmiq proposes waits here for a person
Eight signals, one Stable, Watch or Unstable reading, and a stabilization plan
An in-app assistant that answers questions about your quality data
What the product is meant to do, built from your documents and confirmed by your team
Lets coding agents read findings and write scenarios, with every write logged
Bring external verifiers, such as DeepEval for AI features, into the same run grid
Qarmiq proposes new checks and learnings as it discovers more about your product. Your team reviews those proposals, deciding what becomes part of the product knowledge used in future checks.
Nothing is kept until a person accepts it, and your coding agent can read all of it over MCP.
We’ll set up Qarmiq against your product and walk you through a real verification run, from the behavior being checked to the evidence behind the result.