- The greatest risk in agentic AI may not be individual agents, but millions of agents interacting across shared systems
- Telecom's decades of experience managing complexity may provide a blueprint for governing autonomous AI
- The future challenge is not intelligence, but coordination, interoperability and accountability
As the technology industry races toward agentic AI, most of the discussion has focused on how individual agents will reason, learn and act. Far less attention has been paid to what happens when thousands — or eventually millions — of autonomous agents begin operating simultaneously across networks, clouds, enterprises and critical infrastructure.
That question sits at the heart of what I have called "agentic chaos theory": the idea that the greatest challenge of the AI era may not be creating intelligent agents, but managing the interactions between them.
The irony is that one industry has already spent decades wrestling with a remarkably similar problem.
Telecommunications networks are among the most complex systems ever built. They span millions of devices, thousands of applications, multiple vendors, countless policies and billions of daily interactions. When failures occur, they are rarely caused by a single component malfunctioning. More often, they emerge from the unexpected interactions between systems that are all behaving exactly as designed.
And that is precisely the challenge agentic AI is about to create.
As enterprises deploy autonomous agents across networks, clouds, applications, supply chains and operational systems, the industry is becoming obsessed with intelligence while largely ignoring coordination. Yet history suggests that intelligence alone does not create stability. Governance does. Visibility does. Shared context does.
That is why telecom may have more to teach the AI industry about autonomy than Silicon Valley realizes.
To explore that idea, I spoke with Joe Cumello, SVP and GM of Blue Planet, the network automation and operational intelligence company spun out of Ciena. For years, Blue Planet has helped service providers orchestrate, assure and automate some of the world's most complex operational environments.
As enterprises move toward autonomous AI systems, the disciplines developed inside telecom — visibility, governance, orchestration, assurance and closed-loop automation — may prove just as important as the AI models themselves.
Below is our conversation:
Steve Saunders: Most discussions around agentic AI focus on the behavior of individual agents. Do you agree that the greater long-term challenge may be the interaction between large numbers of autonomous agents operating across shared infrastructure?
Joe Cumello: I do, Steve. The bigger long-term challenge is not one agent making a bad decision. It's thousands of agents all making reasonable decisions at the same time, but with competing objectives.
Each agent will optimize for something different: cost, speed, security, energy use, compliance or customer experience. Those goals can easily conflict with one another. One agent may improve its own outcome while creating problems elsewhere.
Telecom has shown us this for years. Networks often fail not because one component or route is changed but because millions of working components interact in unexpected ways.
That is what concerns me most about agentic AI: emergent behavior. Large numbers of agents could create instability because they lack shared context, shared policy and visibility into the wider impact of their actions.
The real risk isn't rogue AI. It's thousands of autonomous agents pursuing legitimate goals that are incompatible.
Saunders: Something needs to be in charge, right? What is the equivalent of air traffic control for agentic AI?
Cumello: Great question. For me, air traffic control for agentic AI is a network of agents operating under shared decisioning, orchestration and governance frameworks rather than one central AI managing each agent.
To make sound decisions, agents need common context: shared data, policies, objectives, guardrails and validation. But context alone is not enough. Effective oversight requires knowing which agents exist, what authority they have, what they are trying to do, how they interact and when they should be constrained, escalated, or overridden.
In telecom this becomes especially important because agents will operate across network, IT, digital, business, consumer and operations domains. If those agents are not coordinated, they can easily work at cross-purposes.
Autonomous agent management becomes critical. That includes agent catalogs, lifecycle management, security controls, governance guardrails, explainability and traceability. You need to be able to understand what happened, which agent acted and why, so you can quickly course-correct when necessary.
Saunders: What lessons from network automation and assurance are most relevant to the emerging agentic AI era?
Cumello: Telecom has already learned a key lesson that applies directly to agentic AI: you can't automate what you can't see.
Successful automation depends on visibility, context and data integrity. Accurate inventory serves as operational truth. Agents need a clear understanding of resources, services, topology, dependencies, policies and customer impact. Without that context, they're making decisions blindly.
A second lesson is that automation without governance creates risk. Many failures stem from misconfigurations, untracked changes, or automation acting on incomplete information. Agents need guardrails, policy controls, change validation and an understanding of the broader impact of their actions.
The third lesson is that closed-loop automation only works when assurance and execution are connected. Agents must be able to observe, reason, act, validate outcomes and adapt. It has to be a full circle.
The age of viewing autonomy through the lens of a single application is over. Agentic AI and autonomous networks depend on the interplay among applications, governance and data to create business outcomes.
Saunders: Will we move toward open coordination frameworks, or competing proprietary agent ecosystems?
Cumello: Proprietary agent ecosystems will probably emerge first because every major platform provider will want to own the agentic layer.
But business processes do not stay within a single platform. Telecom and enterprise environments depend on many systems working together, including legacy systems.
As agents spread, they will need to collaborate across vendors, platforms and organizations. At that point interoperability becomes less of a design preference and more of a business requirement.
There will be increasing demand for open standards, coordination frameworks and governance models — not because openness is idealistic, but because business scale requires it.
Saunders: How do orchestration, assurance, inventory and operational intelligence evolve in an agentic world?
Cumello: The categories themselves begin to blur.
Inventory evolves from a database-driven view into a knowledge graph. Orchestration moves from workflow automation to translating intent into action. Assurance expands beyond fault and performance management to continuously validate that autonomous decisions are producing the right outcomes.
Operational intelligence becomes the system that coordinates all of these functions, with every autonomous action drawing from shared knowledge and common data.
Historically, inventory, orchestration and assurance existed as separate operational domains. That separation becomes increasingly artificial in an agentic world because autonomous intelligence requires all three simultaneously. It needs contextual knowledge, the ability to act and the ability to validate outcomes.
That's why we're seeing the early stages of OSS evolving into something broader: an operational intelligence layer designed to supervise autonomy itself.
Saunders: Looking five years ahead, what is the industry underestimating most?
Cumello: Governance. Much of today's discussion focuses on model performance. But the harder challenge is decision accountability.
When autonomous agents begin acting across shared systems, who is responsible when one agent affects another? How are conflicts between competing objectives resolved? Who determines whether an outcome was correct? How are autonomous decisions audited? Those questions become far more important than benchmark scores.
The conversation we should be having is not: "How smart can agents become?" It's: "How do we govern autonomous decision-making at scale?" Because eventually intelligence will not be the hard part. Coordination will be.
The future will not be defined by the smartest agent, but by the systems that allow millions of agents to work together safely, predictably and transparently.
And there is a powerful economic reason to get this right. Operators view AI and automation through the lens of enterprise value. Autonomous operations can eliminate redundant applications, reduce integration costs, simplify technology estates and improve efficiency. Those savings ultimately flow directly to the bottom line.
That may be the most compelling reason of all for solving the governance challenge before agentic AI reaches full scale.