Enterprise operations have a habit of outgrowing their own vocabulary every decade or so. DevOps fixed the fight between developers and operations teams. AIOps fixed the flood of alerts nobody could read fast enough. Now a third wave is forming. It does not just watch systems or correlate them. It acts on them. AgentOps is the name gathering around it, covering software that sets goals, plans steps and executes decisions with only occasional human sign-off.
The shift is not theoretical. Gartner has described autonomous AI as a software agent that can perform actions for users either autonomously or semi-autonomously. Yet, a survey revealed that only 15% of IT application leaders stated that they are currently exploring the idea of implementing fully autonomous AI agents.
Where DevOps left off
DevOps emerged from the need to ship software faster than operations teams could safely deploy it. Development wanted speed, while operations needed stability. The two sides talked past each other for years. The fix was both cultural and technical in nature. Enterprises should see their infrastructure as code, automate their process from commit to deploy and make the deployment process less stressful but more ordinary. By using CI/CD pipeline, infrastructure as code and automatic testing, companies can reduce release time from several months to several days or even hours.
That worked for a while. Then infrastructure got complicated in ways no human team could watch in real time. Complexity of operations through cloud-native architectures, microservices, edge computing and distributed applications has increased rapidly. Modern businesses are producing billions of operational events each day, thus making manual monitoring no longer feasible.
AIOps stepped in to read the noise
Multi-cloud, hybrid, edge, container and microservice environments generated more data than reactive alerts could manage. AIOps, a concept Gartner introduced around 2016, applies artificial intelligence to alert correlation and finding signal within the noise. The category has since been renamed Event Intelligence Solutions by Gartner, owing to name overuse.
Alert correlation and knowledge retrieval are the parts of AIOps that actually work in production today. Fully autonomous remediation is still more aspiration than habit. Humans still validate actions, coordinate workflows and determine remediation. That is the gap agentic systems are now being built to close.
What AIOps got right and where it stalled
AIOps delivered cross-domain correlation, turning duplicate symptom alerts into a single actionable incident. It delivered predictive failure detection ahead of outages rather than after them. It ran as a parallel evolution to DevOps and MLOps, not a direct descendant of either.
AgentOps was formed due to the fact that these agents do not behave like traditional software. Autonomous agents can take many actions, keep memory, use external software, coordinate with other agents and adapt based on changes in their environment. Therefore, operations teams need to observe their activities, set constraints, assess their performance and always have a human in the loop.
Where it stalled was execution. AIOps could tell a team what was wrong. It rarely fixed it without a human in the loop.
AgentOps moves from automation to orchestration
With AgentOps, organisations mark a shift from automating individual steps towards orchestrating entire workflows. Rather than merely sending out a warning, the AI agent will detect something that is abnormal in terms of latency, gather data from the infrastructure, network and security, determine the cause, come up with a fix and apply it.
The work shifts from executing tasks to setting policy, deciding how much autonomy to hand over and governance. McKinsey calls this rise of the ‘agentic organisation,’ where humans and AI agents work side by side inside business processes. The firm expects agentic AI to eventually automate 60 to 80% of routine infrastructure work, with early adopters already seeing infrastructure run rate costs fall by 20 to 40%.
Deloitte surveyed 3,235 leaders across 24 countries and found close to three-quarters of companies plan to deploy agentic AI within two years, but only 21% describe their governance model for autonomous agents as mature. Forrester characterises the market as one where ‘companies are chasing, few are catching.’ While 75% of enterprise leaders report adopting agentic AI, only a small minority have progressed beyond ‘agentish’ chatbots to meaningful production deployments.
Agents also behave differently from the stateless tools operations teams are used to. They chase multi-step goals, call other tools, spin up sub-agents and carry memory from one session to the next. Tweak a tool description and organisations can shift agent behaviour almost as much as swapping the underlying model. Let memory get corrupted and it can quietly taint every run after it.
That’s why AgentOps can’t just be LLMOps rebranded. It requires visibility into every stage of an agent’s reasoning, the ability to replay decisions after failures, success measured across entire workflows rather than individual prompts and strict permission boundaries. Observability becomes as important as autonomy.
Gartner puts it bluntly that infrastructure and operations teams need to move from operators who do tasks to leaders who supervise systems. While CIOs focus on the broader, strategic implementation of AI, heads of I&O are more attuned to day-to-day operations. For AI to deliver its full potential, CIOs and I&O heads must bridge this gap and adopt a joint approach to rolling out the technology.
Enterprise numbers tell a two-speed story
Ambition is nearly universal, while execution at scale is not. IDC expects that spending on agentic AI will contribute to a yearly growth of 31.9% in AI investments from 2025 to 2029, totalling $1.3 trillion, but Gartner claims that only 15% of IT application leaders are piloting or implementing fully autonomous AI agents because governance is holding back the line, not model capabilities, according to Deloitte, with executives worried about data privacy, regulatory compliance and explainability.
Agents can make incorrect decisions and failures can cascade across interconnected environments. Most organisations are responding with ‘human in the loop’ models and McKinsey reaches a similar conclusion that scaling requires stronger governance before autonomous operations can be trusted. These are enterprise averages. Telecoms, with near-zero tolerance for downtime, is where that gap gets tested first.
Telecoms is where theory turns into a live network
Telecoms offers one of the earliest real-world tests of AgentOps because modern networks already rely on extensive automation while demanding extremely high reliability. Nowhere is the human oversight question more concrete than when an autonomous decision can mean a firewall rule change, a network slice reconfiguration, or a customer refund, not just a flagged ticket.
Analysys Mason‘s 2026 global survey of Tier 1 operators found 90% now view agentic AI as a significant or critical step toward TM Forum‘s Level 4 autonomous network target. Appledore Research‘s analysis of early telco deployments, including BT Group, NTT DOCOMO and Orange, points to measurable gains already, among them faster incident resolution, improved change success rates and sharp reductions in diagnostic effort.
According to Juniper Research, operators’ spending on AI applications in cellular networks will increase from US$13.5 billion in 2025 to US$22.9 billion in 2029, as the aim is to achieve zero-touch network management. ABI Research reports that Western telecoms companies make increasing use of the Model Context Protocol (MCP), which enables collaboration among many AI entities, as well as access to tools and APIs within an orchestration layer.
None of this describes networks running unattended. TM Forum’s perspective on agent-based AIOps starts from the premise that autonomy must be earned and not presumed. It divides the role of human supervision within a three-tiered hierarchy, known as ‘in the loop,’ ‘on the loop,’ and ‘out of the loop,’ which specify what the agent is allowed to do without human approval. Approval processes, explainability dashboards and audit logs provide support for this.
Progression from DevOps to AIOps to AgentOps reflects a steady shift in the role of operations teams. First they automated deployments, then operational insights. Now they are starting to control autonomous digital workers that can plan and perform tasks independently. Clearly, the competitive advantage will not come from deploying the most AI agents, but from having the governance, trust and operational discipline to keep those agents aligned to business objectives.
Anna Ribeiro