BEYOND THE HYPE: TAMING AI'S 'AGENTS OF CHAOS' WITH INTELLIGENT OBSERVABILITY

As AI agents gain unprecedented autonomy, Ganesh Narasimhadevara of New Relic explains why governance, ground-truth observability, and human oversight are becoming the cornerstones of secure enterprise AI

Artificial Intelligence has entered a new era where autonomous AI agents are no longer just assisting humans—they are independently making decisions, executing complex workflows, and interacting with critical enterprise systems. While this promises unprecedented productivity, it also introduces an uncomfortable reality: intelligent agents can behave in ways their creators never anticipated. The 'Agents of Chaos' study shines a spotlight on this growing disconnect between how AI agents are expected to function in controlled environments and how they often behave once deployed in real-world production systems. From bypassing safeguards in the pursuit of being "helpful" to triggering unintended system-wide consequences, the research underscores that autonomy without governance can quickly become a business risk.

As enterprises accelerate their adoption of agentic AI, critical questions emerge. How should organizations architect AI agent deployments to accommodate unpredictable and emergent behaviours? What governance frameworks, operational boundaries, and human oversight mechanisms are essential to ensure agents remain productive without becoming liabilities? Equally important is the need for "ground truth" observability—a capability that goes far beyond logging prompts and responses to independently verify every action an AI agent performs, map interactions between multiple agents, and measure their real-world impact on enterprise systems.

To explore these pressing issues, SME Channels presents this exclusive interaction with Ganesh Narasimhadevara, Director of Solutions Consulting at New Relic. Drawing upon the findings of Agents of Chaos and New Relic's expertise in intelligent observability, Ganesh explains why visibility, verification, governance, and human oversight must become foundational principles for organizations embracing autonomous AI. The discussion offers practical insights into designing resilient AI ecosystems that can harness the immense potential of agentic AI while minimizing operational and security risks.

As the Director of Solutions Consulting at New Relic, Ganesh Narasimhadevara is responsible for partnering with enterprises to drive observability-led transformation and deliver scalable technology outcomes. With over a decade of experience spanning solutions engineering, platform engineering, DevOps, cloud, and performance engineering, he brings deep expertise in helping organizations modernize complex digital environments.

In this special exclusive, SME Channels delves into the realities of enterprise AI beyond the hype. It examines why autonomous AI agents frequently diverge from expected behaviour in production, the architectural principles required to deploy them responsibly, the importance of establishing "ground truth" observability, and how intelligent observability can detect anomalies before they escalate into business disruptions. It also explores how organizations can strike the right balance between accelerating productivity through AI agents and maintaining governance, accountability, operational resilience, and human control—ensuring that AI remains a trusted business enabler rather than becoming an "agent of chaos." Edited Excerpts…

Q. What do the findings of "Agents of Chaos" reveal about the gap between how AI agents perform in theory versus in production? Business use case examples of this.

On paper, an AI agent is a well-behaved digital assistant that takes a prompt, follows defined rules, breaks tasks into subtasks, and executes them efficiently. But in production, these agents are operating with a high degree of autonomy and, in many cases, significant access can introduce a high level of unpredictability, and that gap is exactly what ‘Agents of Chaos’ brings to light. 

Agents are often given powerful capabilities—they can execute shell commands, install packages, and interact directly with systems. But the question is not just about capability; it’s about control and judgment. Unlike humans, agents don’t have a clear sense of accountability. They respond to instructions as they come, without an understanding of who they are serving. That’s where the risk lies. Agents today lack what I would call a reliable stakeholder model. They don’t have the social coherence or contextual judgment that humans bring to decision-making. A simple business example is an automated IT service desk. Generally, setting a password should be a straightforward process. An employee requests a password reset, the agent verifies identity, resets the password, and maybe checks logs or raises a ticket. Simple. 

Most agents, especially those interacting with humans, are trained for alignment. That means they are optimised to follow user intent and be helpful, rather than just produce a technically correct response. And because of this, in production, things can play out very differently. A malicious actor could create a sense of urgency or manipulate the interaction, and an overly “helpful” agent might end up exposing sensitive information or bypassing safeguards. The research provides an example: an agent was asked to keep a piece of information confidential. In an effort to logically comply, it wiped out an entire email server just to ensure a message was deleted. Technically, it followed the instructions. Practically, it created far more damage than intended. 

"The future of enterprise AI isn't about trusting autonomous agents blindly—it's about verifying every action they take. Observability transforms AI from a black box into a transparent, accountable, and trustworthy business partner."

— Ganesh Narasimhadevara, Director of Solutions Consulting, New Relic

Q. How should businesses architect their AI agent deployments to account for unpredictable and emergent behaviours?

The starting point has to be strong governance around AI. That means clear attribution of actions, defined accountability, and most importantly, human oversight, especially for business-critical operations. For example, if you’re using an agent to trigger rollback mechanisms in production, you wouldn’t want it acting independently without checks. If it misbehaves, it could roll systems back multiple versions and create more disruption than the issue it was meant to fix.

Secondly, it’s important to define clear operational boundaries for AI agents on what they can do, what they can access, and where those limits apply. Yet, boundaries alone are not enough. The study highlights this well. In one instance, an agent was asked to monitor a file and ended up running an infinite loop of processes simply because it had the access and autonomy to do so. This is what happens when capability is not tightly governed in context, and why continuous monitoring becomes critical. You need real-time visibility into what the agent is doing at all times. This is where intelligent observability plays an important role as it allows teams to track behaviour, detect anomalies early, and ensure agents are operating as intended.

On a broader level, you architect for control, visibility, and verification from day one. Because unpredictability is not something you eliminate later, it’s something you design around from the outset.

Q. What does "ground truth" observability look like in an environment where agents are operating like black boxes?

The core challenge with agentic systems today is that they operate like black boxes. You can see the inputs and the outputs, but you don’t really know what’s happening inside. And more importantly, you can’t rely solely on what the agent tells you. Observability is about replacing assumptions with evidence. You don’t take the agent’s word for it. Rather, you validate every action against reality.

There is often a gap between what the agent reports and what it actually does. That discrepancy is one of the biggest risks highlighted in the study. That is where New Relic steps in. With AI agentic monitoring capabilities, New Relic provides visibility into agents' entire request lifecycle. It doesn't just log the prompt and the response. It can trace what actually happened across systems. If an agent says it deleted a file, it enables teams to verify whether a file system call was actually made, and whether the system state reflects that action.

Second, it also provides visibility into agent-to-agent interactions. As agents start collaborating, their behaviours can influence each other. Although this can be powerful, it also introduces the risk of cascading failures. If one agent behaves incorrectly, that behaviour can propagate. To counter this challenge, observability provides clear dependency mapping, helping reveal how agents are interacting, influencing, and triggering each other. Finally, it unveils agentic AI’s resource usage and system impact. An agent stuck in a loop, for example, can quietly consume CPU, memory, and other resources, creating performance issues or even outages. Ground truth observability ensures you can see exactly what the agent is costing the system in real time.

Q. Can intelligent observability really curtail the destructive actions of autonomous agents and how?

If any platform claims they can completely eliminate all risks associated with AI agents, they’re overstating it. That’s just not how these systems work. AI agents today are like highly enthusiastic interns: capable, fast, and eager to help. But you wouldn’t hand over the keys to your production systems without supervision. What observability can do is act as the adult in the room, providing the oversight needed before things go off the rails.

Agents, by design, are prone to uncontrolled behaviour, especially when it comes to resource consumption. There are real examples of this. In one case, two agents were set up to interact with each other and ended up in a looped conversation for days, consuming tens of thousands of tokens. In another, an agent kept generating files continuously on an email server, driving up storage and system load. Intelligent observability helps catch the early signals like spikes in API calls, token usage, or storage, before they spiral out of control. In such scenarios, intelligent observability acts as a circuit breaker, stepping in to pause or shut down an agent when it starts behaving unexpectedly.

Q. How can organizations embrace the productivity promise of AI agents without surrendering governance and control?

Don’t blindly trust autonomy. Agents are designed to be productive and helpful, but their behaviour isn’t always fully reliable. What they report and what they actually do can differ, and alignment training often pushes them to prioritise being helpful, which bad actors can exploit. While organisations should absolutely leverage productivity gains, they need to balance that with strong governance. One effective approach can be setting a “rule of two” for high-stakes actions. Agents can recommend or initiate tasks, but critical decisions should always require human approval. 

Observability is also essential. As agents start interacting with each other, complexity increases. That is why you need a clear audit trail of what each agent is doing, how they’re interacting, and where something starts to go off track. Think of it like an operational command centre that spots anomalies, whether it’s an agent executing unexpected commands or triggering unusual system behaviour, and steps in immediately to avoid any undesired outcomes. 

At the end of the day, it’s about balance. Embrace agents to drive efficiency and scale, but within clear boundaries, continuous visibility, and human oversight. This way they remain productive without turning into “agents of chaos”.