QAwerk founder: ‘The real danger is systems drifting into unsafe behaviour’
- Michiel Willems
- June 10, 2026
Konstantin Klyagin, founder of QAwerk
As financial institutions accelerate the move from standalone AI tools to autonomous, agentic systems, QA and software testing teams are being pushed into unfamiliar territory: validating not just outputs, but behaviour across entire workflows.
The shift is already visible in production environments. As agentic AI moves from experimentation into live banking systems, AI agents are increasingly interacting with each other, passing signals, triggering actions and shaping outcomes across payments, fraud, trading and customer-facing processes.
Konstantin Klyagin, founder of QAwerk, says the implications for QA teams are immediate and structural.
“Once AI agents begin interacting, the unit of risk is no longer a single output. It is the chain of decisions, handoffs, and side effects that unfold across systems over time,” Klyagin told QA Financial.
Traditional QA approaches in banking have focused on validating whether a model produces the correct output for a given input. But as AI systems begin to act autonomously and interact with other systems, that approach is breaking down.
“In banking, that means QA teams cannot stop at asking whether one model produced the correct answer in one prompt,” Klyagin explained.
“They need to test whether a sequence of AI-driven actions remains safe, consistent, and controllable once it touches payment rails, fraud engines, case management tools, or customer-facing workflows.”
“In banking, QA teams cannot stop at asking whether one model produced the correct answer in one prompt.”
– Konstantin Klyagin
In practice, this is shifting testing toward orchestration behaviour, where the focus moves to how systems behave in combination rather than isolation.
“You have to validate whether agents preserve context correctly, escalate at the right moments, respect permissions, avoid feedback loops, and fail safely when one component misfires,” he pointed out.
The risk, Klyagin added, is that “a model can look accurate in isolation and still create operational risk once it starts triggering downstream actions in production.”
Invisible failure modes
As multiple AI systems interact, new classes of risk are emerging, many of which are difficult to detect using traditional testing frameworks.
“The biggest new risk is compounding error,” Klyagin said. “One agent makes a flawed assumption, passes it to another system as if it were reliable, and that second system acts on it with real consequences.”
In a banking context, those consequences can be immediate and material.
Konstantin Klyagin
“In financial services, that can mean a payment routed incorrectly, a fraud case escalated or dismissed without enough grounds, or a trading-related workflow responding to the wrong signal.”
Alongside compounding error, firms are also facing growing challenges around attribution and control.
“When multiple agents operate across the same workflow, it becomes harder to see who made which decision, why that decision was made, and whether the action stayed within policy,” he shared.
This creates not just governance risk, but operational fragility. Klyagin warned that “you can get silent failures, duplicated actions, inconsistent customer treatment, or decisions that are individually rational but collectively wrong.”
“The real danger is not a single agent failing, but multiple agents appearing to work while the system as a whole drifts into unsafe behaviour,” he added.
Testing flows, not snapshots
For QA teams, the shift is forcing a rethink of how testing is designed and executed in production-like environments.
“It means QA has to observe flows, not snapshots,” Klyagin said. “Instead of validating one answer against one expected result, teams need to run multi-step scenarios and watch how the system behaves across repeated interactions, changing inputs, and real operational constraints.”
In banking workflows, this can involve multiple agents acting in sequence across different systems.
“For a bank, that could mean testing what happens when an AI fraud agent flags a transaction, a second agent requests more evidence, a third agent updates a case, and a customer-facing system sends a message, all within the same chain.”
To support this, QA teams need stronger infrastructure around observability and traceability.
“You need to see what each system received, what it inferred, what action it took, and what happened next,” Klyagin explained.
Monitoring, he continued, should focus on “drift, escalation quality, exception handling, and policy adherence over time.”
“QA has to observe flows, not snapshots. Teams need to run multi-step scenarios and watch how the system behaves.”
– Konstantin Klyagin
As systems become more complex, AI is increasingly being used to test AI, particularly to expand coverage and surface edge cases.
“It is promising, but I would not call it fully mature,” Klyagin stressed. “Today, AI can already help QA teams generate test cases, explore edge scenarios faster, review large response sets, and identify patterns that would take humans much longer to surface.”
However, he is clear that this does not remove the need for human oversight. “In a bank, you cannot rely on one opaque system to certify another opaque system.”
“AI can help generate hypotheses, simulate adversarial behaviour, and spot anomalies, but humans still need to define risk thresholds, review outcomes, and decide what is acceptable from a compliance and business standpoint.”
Explainability
From a governance and regulatory perspective, the rise of agentic systems is also redefining what explainability means in practice.
“They need to treat explainability as a property of the workflow, not just the model,” Klyagin shared.
Konstantin Klyagin
“If several agents are involved, regulators will still expect the institution to show who did what, on what basis, under which rules, and with what controls.”
That requires a much deeper level of auditability than many current systems provide.
“That means preserving full decision traces, documenting escalation logic, logging data access and tool usage, and making sure there are clear human override points where needed,” he said.
Klyagin added that architectural clarity is critical. “If responsibilities between agents are vague, explainability breaks down immediately.”
He elaborated: “You cannot defend an AI-driven process by saying the system was too complex to unpack. In regulated sectors, that answer is unacceptable.”
Closing the gap
As agentic AI moves toward wider deployment, Klyagin pointed to three areas where QA teams need to evolve quickly: observability, test design, and governance integration.
“The first gap is observability,” he said. “Most teams still do not have enough visibility into cross-agent behaviour.”
Without it, “QA becomes guesswork,” he argued.
The second gap is test design, as traditional frameworks struggle with non-deterministic systems. “Agentic systems require teams to test ambiguity, changing context, and long interaction chains.”
Finally, governance and testing need to be more tightly integrated. “Financial services teams need controls that connect testing, compliance, and production monitoring instead of treating them as separate layers.”
Underlying all of this is a more fundamental shift in how QA defines correctness. “Traditional, rigid automation frameworks simply will not work for non-deterministic systems,” Klyagin stated.
“You have to evolve your assertions to validate meaning rather than just matching characters,” he explained.
In practice, that means moving beyond exact-match validation toward semantic and structural checks, an approach that is already being applied in real-world AI testing environments.
“For financial institutions, this is the reality of modern QA,” he stressed, concluding that “we have to stop testing for predictable text and start automating for accurate intent, factual coherence, and structural integrity.”