A conversation with Gastón Marichal, QA Manager at QAlified, on why the hardest part of agentic QA has nothing to do with the technology.
In the first installment of this series, we heard from Huy Tieu, a Katalon product manager building an AI quality companion from the inside. This time we wanted the other side of the table: someone who has to make agentic QA work for real client teams, on real delivery timelines, with real consequences if it goes wrong.
Gastón Marichal is QA Manager at QAlified, a testing and quality assurance consultancy based in Montevideo, Uruguay, and a Katalon partner since 2022. He has spent 12 years in testing, and none of it was the plan. "I often say that I learned what testing was in a single class during a course called Software Engineering," he says. "Testing was presented as the final stage before production, following requirements, design, and development. That one session was where I first learned the terminology and theory behind testing." Everything after that, he learned on the job, with a mentor who taught him what he calls "this beautiful art of testing."
That hands-on, trial-and-error instinct shows up throughout the conversation. Gastón doesn't reach for frameworks first. He reaches for what he's watched work, and what he's watched fail, on real teams with real clients watching the results.
Ask Gastón to define agentic QA and he starts with what it isn't. "For me, agentic QA is the evolution of the automation we used to do before. We now have the ability to delegate many tedious, repetitive, and even boring tasks to a powerful tool such as AI agents," he says. "But what I think people are getting wrong is the definition itself. It's not black magic or a superpower. AI is phenomenal, but it is still software that produces results based on variables, trillions of variables and parameters in this case."
That framing matters more than it sounds. A lot of the anxiety and hype around agentic QA comes from treating AI agents as something categorically different from the software testers have always worked with. Gastón's view keeps it grounded: it's still software, it still has failure modes, and testers are the ones best equipped to find them. "As testers, more than anyone, we know how things can go wrong," he says. "So we need to be objective, understand the limitations and risks of AI, and be more critical of its results. AI is incredibly powerful, but power without control is not very useful."
He's also refreshingly honest about where the line sits between AI-assisted testing and something more agentic. "Honestly, I don't know exactly how to define them separately, or if one includes the other," he admits. "But I think AI-assisted testing is more broad and general, and agentic QA is one example of how we can use AI to help us with testing tasks." It's a more modest claim than the confident category lines vendors often draw, and arguably a more useful one for QA managers trying to cut through the noise.
Gastón has used Katalon since the early Studio-only days, which gives him a long view of where fully scripted automation started to strain. "As software reduced its development and release cycles, with constant changes that broke our test scripts and everything else, we started to feel that this 100% code-based automation, without any help for maintenance or quick updates, was accumulating a lot of work and effort for us," he says.
Where his perspective diverges from a product builder's is in what he blames for the industry's automation coverage ceiling. It's not primarily a tooling gap. "I think the clearest limitation today is the desire to constantly release to production, to have everything developed and deployed in a matter of minutes, and the constant rush to deliver things faster," he says. "If we don't have a project structure, a team that sees automation as a strategic need, and not just something for the tester to take care of, and a methodology that benefits automation, then achieving good automated test coverage is difficult."
In other words: better AI doesn't fix a coverage ceiling caused by treating automation as an afterthought. That's an organizational problem before it's a technical one, and QA managers run into it long before they ever get to evaluate an agent.
Trust is the recurring theme of this series, and Gastón's version of it is shaped by client delivery, not just internal team buy-in. "I think a controlled and gradual adoption, supported by some kind of governance or monitoring framework, helps guide those who are enthusiastic about AI while also exposing results and presenting success stories to those who are more skeptical," he says. With clients, the same instinct applies outward: "Being transparent about how AI is being used, for what reasons, and what the benefits would be for them is fundamental to avoid conflicts."
He's also candid about getting it wrong. "Yes, several times, I would say," he says, when asked if AI testing has ever given him a false sense of coverage. "But it has all been part of the learning process for me: understanding what works, how to ask AI for tasks, and when it is worth taking a deeper look at the outputs. Hallucinations and false assumptions are still present in the models, and that's where we face a big risk if there isn't a person, someone actually thinking, behind them to verify the results."
It's a plainer version of the same conclusion Huy Tieu reached from the product side in the first installment of this series: the agent isn't the safeguard. The human oversight and governance built around it is.
Gastón leads teams across automation, performance, and AI testing, and the shift looks different depending on which discipline you're standing in. "In terms of automation and performance testing, what has changed the most are the strategies and mechanisms used to do the work: tools, activities, timelines," he says. "And in terms of AI testing, new roles and specializations have been created in this field." That tracks with what's happening more broadly across how QA teams are restructuring around agentic work.
He's sharper about the anti-pattern he sees most often: applying AI everywhere, regardless of whether it solves anything. "The excessive use of AI, without any apparent control, and the forced use of AI in tasks or activities where its value is not really clear," he says. "AI is capable of helping in almost any area, but it depends on people to apply it where the real pain points are, or where we see that it can provide the greatest value." It's a useful check for any team currently under pressure to look "AI-forward" for its own sake.
Developers are shipping faster with AI-assisted coding, and Gastón's answer to what that means for QA isn't defensive. It's an invitation to use the same leverage. "I think that today we have the opportunity to use the same tools and development capabilities that are available to developers," he says.
"We need to use that to the tester's advantage, stay ahead of the problems and challenges of each development cycle, and keep learning and updating our skills to stay on par with development. That allows us to provide the added value that testers bring: the critical and objective perspective that has always characterized us."
Keeping validation moving at that speed without burning out the team, in his view, comes down to prioritization discipline rather than working harder. "More than ever, we need to have a clear testing strategy and a solid mechanism for prioritization and risk analysis," he says.
And he's blunt about what the goal isn't: "The goal of the tester should not be to execute tests in a mechanized way, with or without AI, it doesn't matter, just to achieve a green dashboard. Instead, the tester should provide a critical evaluation, focused on a software product that we increasingly know less about in terms of how it was implemented, while never losing sight of the end users."
When asked what actually stops teams from making this shift, Gastón doesn't point to budget or infrastructure.
It's the sharpest observation in the conversation, and it lands because it isn't really about AI. It's about a profession where thoroughness and manual effort have long been proxies for rigor, now being asked to redefine what rigor looks like when the effort moves elsewhere.
Gastón's practical advice for a QA manager wanting to start is deliberately small. "They should start with two tasks," he says. "The first is researching and learning what people who are doing agentic testing are doing today. The second is to test, test, and fail as many times as necessary until they find something that is useful for them and that they can replicate in their teams or projects."
He adds a team-management move worth borrowing directly: identify who's enthusiastic and who's skeptical, and use both. "Identifying promoters and detractors within the teams also helps: one brings synergy and the other brings caution, and working with them on how to approach agentic testing." Promoters build the first proof points. Detractors stress-test whether those proof points actually hold up. Neither group should be managed out of the process.
Gastón Marichal has spent 12 years learning testing by doing it, and his read on agentic QA follows the same instinct: don't wait for certainty, run small experiments, and let the results, not the pitch deck, do the convincing. The tools are more accessible than they've ever been. What's still catching up, in his experience, is the culture around using them.