17-06-2025
The Future Of AIOps Is Many Agents Working Together
Karthik Sj, General Manager, AI at LogicMonitor. Built & Scaled multiple 0-1 AI products across public, PE and VC backed companies.
If it feels like everyone is suddenly selling "AI agents," you're not imagining it. They're everywhere, pitched as automation 2.0: fewer tickets, faster resolution, more time for strategic work, replace the help desk, reinvent the NOC and free your team from toil.
But behind the confident marketing is an open secret: There's no consensus on what an AI agent actually is.
Definitions vary wildly depending on who's talking—and what they're selling.
Some define agents as tools that use large language models. Others frame them as autonomous systems capable of observing, deciding and acting independently. In some documentation, "agent" is just a new label for a chatbot with plug-ins. In others, it implies a complex, goal-driven process manager.
Even the terminology is unstable. Some use "agent" and "assistant" interchangeably. Others draw a hard line between them. A few acknowledge the ambiguity outright, admitting that the term can describe anything from rule-based workflows to reasoning systems.
The term "agent" is being stretched past the point of meaning, making it harder for teams to evaluate real capabilities and turning every vendor demo into a guessing game.
If "agent" is going to be the core metaphor for how AI works in the enterprise, it's worth getting clear not just what these systems do but how they do it, how they interact and where the limits actually are.
So let's reset.
At the most practical level, an AI agent is a system that can observe, reason, act and adapt in pursuit of a goal with as little human intervention as possible.
That's it. The key isn't that it's "smart." It's that it has agency.
The term "agent" comes from "agency": the capacity to take meaningful action toward a goal. Not to wait but to initiate. To sense the environment, decide and do. That's what gives agents their power.
Most AI solutions marketed as 'agents' today fail that definition. They don't observe; they wait for prompts. They don't reason; they pattern-match. They don't act; they suggest. They don't adapt; they repeat.
Strip away the marketing, and most are just prompt-driven wrappers around existing functionality.
And that's fine—as long as we're honest about it. What matters is what the system can actually do.
The next leap forward in AI for ITOps isn't building smarter standalone agents; it's engineering modular systems within which multiple specialized agents solve problems as a unit.
Why? Because incidents in production environments rarely follow a script. They're layered. They span telemetry, infrastructure, services and human teams. They require context, judgment, escalation paths and institutional memory. A single agent can't handle that complexity.
Take a real-world example: a business-critical application is running slowly.
• A correlation agent clusters related alerts across infrastructure layers—VMs, databases and network components—into a single incident.
• A diagnostic agent identifies the likely root cause: a dependency bottleneck between the application and its backend services.
• A retrieval agent checks for a runbook, which includes manual steps, required approvals, system integrations and references to automation playbooks.
• A remediation agent surfaces the most relevant playbook (e.g., reverse a config change, increase disk space, restart a failing collector), along with alternatives. The user confirms or selects the right one.
• The orchestrating agent executes the playbook using automation tools. If no playbook exists, the system flags the issue for manual resolution and captures the pattern for future automation.
• A verification agent checks whether the system has recovered and performance has normalized.
• A summarization agent compiles the full incident timeline and updates the ticketing system and internal documentation.
Each agent plays a distinct role—detection, reasoning, action, documentation—but none act alone. It's the orchestrator that routes requests, maintains context and sequences multiagent workflows based on the user's intent and the system state.
With an agentic system, by the time you're alerted to a P1 incident, multiple agents may already be working the case: clustering alerts, identifying root causes, fetching past incident data and, in some cases, starting remediation. You're not beginning with a blank screen. You're stepping into an investigation already in motion.
This is all to say that the obsession with 'smarter' agents is a distraction.
You don't need a genius agent that can do everything. You need a suite of agents that know their role, play well and communicate with others.
That means:
• Specialization, not generalization
• Communication protocols, not black boxes
• Predictable behavior, not vague promises of 'autonomy'
If that sounds familiar, it's because we've already solved this problem in the human world. It's how high-functioning teams work.
When you're evaluating whether an 'AI agent' is real or just repackaged automation, ask:
• Can your agents communicate with each other or only with humans?
Good: 'Agents share structured context and trigger each other to take action based on system state.'
Bad: 'Our agent sends alerts to Slack.'
• How is work divided among agents?
Good: 'We have agents for monitoring, diagnostics, remediation and notifications.'
Bad: 'Our agent can handle any use case.'
• What happens when things don't go according to plan?
Good: 'Agents can pull in other agents or escalate dynamically.'
Bad: 'Our agent was trained on millions of incidents.'
If the vendor talks more about how smart the agent is than about how the system works, it's okay to be skeptical.
If you're confused by the flood of AI agent announcements, rest assured you are not alone.
The industry hasn't agreed on what 'agent' even means yet. Rapid innovation has led to rushed narratives, vague demos and a race to sound future-ready without doing the hard system design work.
But this confusion is also a gift. It creates space to ask better questions—about what's real, what's useful and what's needed.
The next phase of AI won't be won by whoever builds the smartest-sounding agent. It'll be led by teams who design agentic systems that can coordinate and specialize.
That's the agentic AI worth building.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
