Photo with Robert Hommes Credit: Robert Hommes TL;DRConventional observability tools were built for deterministic systems and rely on known failure patterns like HTTP error codes. AI agents can complete a task while producing the wrong outcome, and existing monitoring records a success while the business experiences a failure. Moyai founder Robert Hommes argues for anomaly-first detection: find what is different, then determine whether it is wrong, rather than chasing each new failure with another rule. The rapid adoption of AI agents is changing the architecture of how organizations operate.
Agents can interpret information, make decisions, interact with enterprise systems, and execute tasks with a degree of autonomy that would previously have required human involvement. That autonomy creates significant opportunities for efficiency, yet it also introduces a fundamental challenge that conventional approaches to monitoring were never designed to address. For Robert Hommes, founder of Moyai, the central question is no longer simply whether an AI agent can complete a task. It is whether an organization can reliably determine that the task was completed correctly. “The most dangerous example of an AI agent is one that successfully completes a task while actually producing the wrong outcome,” Hommes says. “Agents are different from traditional AI that we know from chatbots, because they have tools, and those tools are usually connected to internal business systems.
Any incorrect use of those systems that is not properly registered or communicated can create a significant difference between what you believe is happening and what is actually happening.” That distinction becomes particularly important when agents have the authority to make decisions or initiate actions with financial or operational consequences. Hommes explains, “Imagine a procurement agent instructed to purchase a specific type of coffee bean. The agent communicates successfully with the procurement system, yet consistently submits the wrong product parameter. At the same time, it checks inventory or purchasing records against a different product category.
From the perspective of conventional infrastructure monitoring, the system may be functioning normally. From the business perspective, however, the agent is steadily producing the wrong outcome.” The discrepancy can become significant before it is recognized. A comparable problem could arise in airline customer service. An agent may tell a stranded passenger that a flight has been rebooked, creating the expectation that the reservation is confirmed.
If the underlying transaction never succeeded, the customer may discover the problem only upon arriving at the airport. These scenarios illustrate what Hommes considers an emerging category of failure. The problem is not necessarily that a system generates an explicit error. The problem is that the state represented by the agent’s trace, the organization’s monitoring systems, and the business’s actual state can diverge.
Traditional observability provides powerful mechanisms for identifying explicit technical failures. HTTP status codes offer a straightforward example. A 400-level response signals a problem with a request, while a 500-level response indicates a server-side problem. A 200 response generally indicates that a request was processed successfully.
AI agents can challenge that assumption. Hommes says, “We do not have an error code that says, ‘I reached the endpoint, I queried it with the wrong parameter, and I received something different from what I needed.’ Technically, nothing is failing, but it is not working.” An agent can therefore make a valid request, receive a valid response, and still make an invalid decision. The infrastructure records a successful interaction while the business experiences failure. This creates a reliability problem that cannot be resolved simply by collecting more conventional technical signals.
Organizations have understandably responded by introducing additional safeguards. Human-in-the-loop architectures can require employees to approve consequential actions. Enterprise systems can monitor spending, transaction volumes, and other indicators of material impact. Mature organizations can also establish controls that prevent an agent from moving beyond clearly defined boundaries.
Hommes considers these mechanisms valuable, particularly because they provide layers of protection around autonomous systems. Yet they do not eliminate the underlying challenge. Monitoring for material impact often means waiting until the consequences have become sufficiently large to trigger detection. “Once you have material impact, you will see it, but you are already too late,” Hommes says. “You already have the impact. It had to get worse before you noticed it.” The deeper issue concerns how organizations approach unknown failure modes.
Conventional observability tends to be built around known patterns. When a particular failure occurs, teams identify the pattern and develop a rule capable of detecting it in the future. That methodology can work effectively in deterministic environments where system behavior is relatively predictable. AI agents introduce greater variability.
Their decisions can depend on context, inputs, tool interactions, and interpretations that are difficult to enumerate in advance. An organization can continually add rules for previously observed failures while remaining exposed to behaviors it has never encountered. Hommes describes that process as “whack-a-mole.” Each new failure produces another rule, while the next unexpected failure remains outside the existing framework. “If you are checking a non-deterministic system with rules, you are doing whack-a-mole,” he says. “You are always trying to find the next problem, and you are always going to be a little late.” The alternative, in Hommes’ view, is to change the starting point for detection. Rather than attempting to define every possible way an agent could fail, organizations can first identify behavior that is unusual and then determine whether that deviation represents a genuine problem. “If we want to find unknown unknowns or these kinds of failure patterns, the best place to look is what is different,” Hommes says. “Most systems have a certain utility.
Most of what the system does is good. So if we first look at what is different and then determine whether what is different is wrong, we have much better detection dynamics.” This represents a broader shift in how organizations should think about observability. Hommes argues that the monitoring models developed for deterministic systems cannot simply be transferred to autonomous agents without reconsideration. “Software reliability and observability are very closely tied to deterministic systems and to the way we operated in the past,” he says. “As we change the way we operate because of the functionality agents provide, we have to ask whether the systems we used in the past can monitor these new systems. My view is that they cannot, because agents fail in different ways, including ways we did not know about and ways those monitoring systems were never developed to detect.” For Hommes, the answer lies in treating AI-agent reliability as a distinct discipline and ultimately as a new product category.
Continuous behavioral analysis can complement existing observability, human oversight, and system-level controls by identifying deviations before they become material consequences. The objective is not to eliminate every possible anomaly. It is to make meaningful anomalies visible early enough for organizations to understand and address them.
