IA Agêntica
Agentic operations: are we ready for AI autonomy?
As artificial intelligence agents stop merely recommending and begin to decide and act, context, governance, and security become as important as the intelligence itself.
In the latest articles in this series, I have been discussing a question I consider central to the future of IT operations: artificial intelligence agents can accumulate extraordinary technical knowledge and still struggle to understand the specific environment in which they need to act. We explored this by showing when an agent masters a tool but does not understand the incident and how the next generation of AIOps relies on operational context platforms.
An agent can know networks, cloud, Kubernetes, databases, or practically any technology in depth. It can consult documentation, analyze large volumes of information, and combine knowledge that would normally be spread across different specialists.
But that does not mean it knows a given company’s architecture, the dependencies between its components, the changes made recently, the history of incidents, the affected services, or the relationship between a technical problem and its impact on the business.
This difference between specialized knowledge and operational context would already be important if we were talking only about agents used to support human specialists.
But a new question begins to arise as we move toward so-called agentic AI.
What happens when these agents stop merely analyzing and recommending and also begin to decide and act on the environment?
Recently, during Telco Transformation Latam, in Rio de Janeiro, I had the opportunity to follow a panel dedicated precisely to the evolution of agentic operations.
More than the possibilities of the technology, what caught my attention was the number of challenges that begin to appear when the discussion leaves experimental cases and approaches real operating environments:
- Systemic view.
- Integration across different domains.
- Legacy systems.
- Governance.
- Guardrails.
- Security.
- Human oversight.
- Orchestration of multiple agents.
- Costs.
- Changes in processes and in team culture.
These are especially relevant questions for Telecom, but they are not limited to that sector. They begin to arise in any critical environment in which we intend to allow artificial intelligence systems to move from recommendation to execution.
And all of them seem to point to a question that comes before the adoption of the technology itself:
Are we preparing our operations to receive autonomous agents as fast as we are increasing the capability of those agents?
Perhaps this is one of the most important questions of the next stage of artificial intelligence adoption in operations.
From automation to autonomy
Automation is not new in IT and Telecom operations.
For decades we have used scripts, rules, workflows, and specialized platforms to perform tasks that previously depended on human intervention. Over time, these automations became increasingly sophisticated and took over important parts of operational processes.
But there is a fundamental characteristic in most of them: someone defined in advance the path to be followed.
A condition is identified, a rule is triggered, and a known sequence of actions is executed. Even when the flow has several branches, there is a previously established logic determining the expected behavior.
Agentic AI begins to change this dynamic.
An agent can receive a goal, observe the environment, query different sources, interpret information, formulate hypotheses, choose tools, decide which steps to execute, and adjust its strategy according to the results found.
This capability opens enormous possibilities for operations.
Imagine an incident that today requires a specialist to consult different platforms, compare alarms, analyze logs, check recent changes, search documentation, identify dependencies, formulate hypotheses, and only then decide what to do.
An agent can potentially take part in several of those steps.
The problem is that there is an important difference between helping someone decide and being given authority to execute the decision.
While AI produces a recommendation, there is still an opportunity for a specialist to evaluate the information, consider aspects that may not have been identified, and decide whether or not that recommendation should be executed.
When the agent starts acting directly on the environment, that last control point may disappear.
In a critical operation, the consequences of this change are significant.
An incorrect interpretation, an outdated piece of information, or a poorly established causal relationship may stop producing only a mistaken recommendation and start producing a mistaken action.
And the greater the autonomy, the greater the impact of that action can be.
That is why the discussion about agentic AI begins to approach themes that traditionally did not sit at the center of the debate about artificial intelligence: governance, permissions, traceability, security, and control.
The journey to Autonomous Networks
In the Telecom sector, this discussion is also advancing within the Autonomous Networks journey developed by TM Forum.
The model establishes progressive levels of autonomy, from L0 to L5. Level 4 represents a particularly important change because it involves more complex environments, including cross-domain ones, in which systems begin to analyze situations and make decisions using predictive mechanisms and closed loops oriented to services and customer experience.
TM Forum itself describes reaching L4 as a shift from traditional automation processes defined by humans to an effective form of autonomous decision making.
This advance is no longer just conceptual. Operators have been declaring and validating L4 in specific domains and use cases, while recent initiatives experiment with agentic AI, orchestrated reasoning, and collaboration across different domains.
But there is a question that comes before autonomy that I consider even more important.
To decide correctly, the agent needs to understand what it is deciding about.
And here we return to the problem of operational context.
The more agents, the greater the need for a shared view
Imagine an agent specialized in access networks. Another in transport. A third in core. Other agents may be associated with cloud, applications, security, or customer experience.
Each one may have excellent technical knowledge and perform its tasks very well within its own domain. That does not mean that, together, they understand the operation.
An incident rarely respects the organizational architecture we built to observe it. A failure may start in a given component, produce symptoms in another domain, affect applications, compromise a service, and finally show up to the organization as a customer complaint.
If each agent observes only its own part of the environment, we may reproduce with artificial intelligence a problem we already know very well in today’s operations: silos.
Today we have silos of tools, data, and teams. Tomorrow we may also have silos of intelligence.
This possibility produces an interesting paradox:
This view does not arise simply because the agents are intelligent.
It depends on information that is normally scattered across many sources: events, metrics, logs, topology, inventory, changes, tickets, dependencies, operational history, service data, and customer information.
More importantly, it depends on the relationships between that information.
An isolated alarm says that something happened.
Context helps explain where it happened, what depends on that component, which other events may be related, what changed recently, and what the impact on services may be.
This is a known problem for human specialists and becomes even more important when we start delegating decisions to machines.
Not by chance, one of the most recent discussions at TM Forum itself about the evolution to Level 4 is related to the concept of the knowledge plane: a layer capable of turning fragmented data into contextualized, reusable knowledge that agents can understand, use to reason, and act on. The goal is precisely to enable cross-domain coordination and more explainable decisions.
Today’s operations were not built for autonomous agents
There is, however, another obstacle. Most of today’s operations were not built imagining that autonomous agents would need to interact with their systems.
The architectures were formed over years or decades. Different vendors use distinct data models. Some platforms have modern APIs; others have limited integrations. Important information may be in structured databases, documents, tickets, or simply in the knowledge accumulated by the teams.
Processes were also designed for people and deterministic automations.
Placing agents on top of this architecture does not eliminate these limitations. In some cases, it may only expose them more clearly.
Before thinking about autonomy, therefore, there are foundational actions: organizing data, integrating platforms, providing adequate interfaces, reviewing processes, defining policies, and building mechanisms that allow agents to access reliable information.
There is also an economic dimension. Agents able to indiscriminately query large volumes of data, trigger models repeatedly, and run long chains of reasoning may be technically interesting and economically inefficient. Managing computational consumption and tokens becomes part of the solution’s architecture itself.
And there is a human dimension.
Adopting agents changes responsibilities, procedures, and skills within teams. Specialists who previously performed certain activities may start to supervise decisions, handle exceptions, define policies, or analyze situations the agents could not resolve.
That is why it may be a mistake to imagine the evolution of agentic AI as a simple replacement of human work. It looks much more like a progressive redistribution of responsibilities between people and systems.
How much autonomy should we hand over?
The answer will probably not be the same for all processes.
An interesting alternative is to think of autonomy as something that needs to be earned progressively.
Initially, an agent may only observe. Then it can formulate diagnoses and recommend actions. In a next stage, it can execute certain tasks with human approval.
As confidence and maturity increase, it can receive autonomy for known, repetitive, lower-risk situations, while higher-impact decisions continue to be escalated to specialists.
It is an evolution similar to the one we naturally apply to people.
A professional who has just joined a critical operation normally does not immediately receive the same autonomy as someone with years of experience. Responsibilities are expanded as knowledge, experience, and trust increase.
With agents, we may need to adopt a similar principle.
The autonomy of agents should not grow faster than our ability to understand, control, and audit their decisions.
This evolution also helps us better understand what the role of human oversight should be in agentic operations.
A concept frequently used in this context is human-in-the-loop, in which certain AI decisions or actions continue to depend on the participation, validation, or approval of a person.
The discussion, therefore, does not need to be reduced to choosing between human-controlled operations or fully autonomous operations. Different decisions may require different levels of oversight.
Reversible, known, low-impact actions can receive greater autonomy. Unprecedented, ambiguous situations, or those capable of producing significant impacts, may require human validation.
The level of autonomy thus also depends on risk, criticality, and trust.
Guardrails become part of the operational architecture
In this scenario, guardrails stop being just a recommendation and become part of the operational architecture.
The agent needs to know not only what to do, but also what it is authorized to do.
Which systems can it access? Which actions can it execute? Under which conditions? Up to what level of impact? How will an exception be handled? How will we later reconstruct the decisions made? How will we interrupt an inadequate sequence?
These concerns already appear formally in TM Forum’s work. Its security guide for Agentic AI in Autonomous Networks deals with risks related to reasoning, memory, tool use, authentication, human interaction, and protocols among multiple agents.
This shows how the discussion is evolving.
When AI only produces information, we mainly evaluate the quality of the answer. When it receives the ability to act, we need to evaluate the entire system around it.
From infrastructure to service and customer experience
There is one last change that I consider fundamental.
For a long time, operations were structured mainly around infrastructure. We monitor equipment, interfaces, links, servers, applications, and a growing number of technical indicators.
But the customer does not buy any of these components in isolation. They buy a service.
A network can show thousands of indicators within expected parameters and still a certain group of customers may be experiencing degradation.
Likewise, dozens of technically different alarms may represent only distinct manifestations of the same failure affecting a single service.
If we intend to move toward more autonomous operations, agents will progressively need to understand not only the infrastructure, but also the relationships between infrastructure, services, and customer experience.
This change of perspective already appears in the most advanced Autonomous Networks initiatives, which combine collaboration between agents and different domains precisely to move from network-centric operations to experience-oriented operations.
At bottom, all these questions converge on the same conclusion.
The development of agents is advancing rapidly. But an agentic operation depends on much more than agents.
It depends on prepared data, integrations, processes, governance, security, guardrails, oversight, interoperability, and, above all, context.
Before building an autonomous operation, we need to build an operation that can be understood.
From vision to practice: ARGUS
The evolution toward more autonomous operations requires more than agents capable of executing tasks. They need to understand the environment in which they act, have access to consistent information, and operate within limits defined by the organization.
This need is directly related to the work we have been developing with ARGUS. The platform started with AIOps capabilities, connecting monitoring, observability, and operations management tools to correlate events, reduce noise, and build a shared operational context.
With the evolution of the proposal to AgentOps, complemented by AI Gateway, ARGUS now also addresses the operation of AI agents themselves: their orchestration, observability, policy enforcement, and tracking of executed actions. The goal is to combine an understanding of the environment with mechanisms that help keep agents under control as they gain autonomy.
After all, it is not enough for agents to be able to act. We need to ensure that they understand the operation and that their actions can be tracked and governed.
In the next article in the series, I go deeper into why autonomous networks do not start with agents, but with an architecture prepared for autonomy.
Learn about ARGUS and see how it works in practice.
About the author
Paulo Florêncio de Menezes is an entrepreneur, partner, and co-founder of companies such as Target Solutions, BiPTT, and ARGUS.
He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).
In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.