A few years ago, I could describe how a production incident would unfold before it happened. When something changed, a signal degraded. A team, sometimes a large one, pulled in from three time zones, followed the trail backward until we found the commit, the config, or the dependency that broke. The work was hard, but the logic held.

That logic is shifting. It isn’t that our systems have become sloppier, but our processes are struggling to keep pace with the reality of modern architecture. We have moved toward systems designed for self-directed operational reasoning and execution across distributed environments — systems that are built to adapt in real time rather than follow the static, predictable routes of the past.

Agents handed a goal will find their own route to it, and the route they pick today may not be the one they pick tomorrow. That autonomy is the whole point of agentic AI. It’s also the end of an assumption we stopped questioning long ago: that understanding a system once means understanding it for the future — as long as we don’t change anything.

To see how teams are adjusting, we surveyed 2,050 ITOps and engineering professionals across ten countries and sixteen industries for “State of Observability 2026: Restoring Predictability to the Agentic Era.”

We found two distinct groups. Most organizations are still adapting their practice to agentic AI, while a smaller group has rebuilt theirs around it.

The arrival of unpredictability  

Operations have been getting less predictable for a decade. Ephemeral infrastructure, third-party sprawl, and release velocity all erode our ability to reason about systems in a straight line.

But that erosion was accidental. Nobody designed microservices to be inscrutable; it was a side effect we managed as best we could.

AI is different. Non-determinism isn’t a flaw we’re trying to engineer out. It’s the mechanism by which agents reason, adapt, and act without a human writing the next step. Survey respondents feel the weight of that: 83% are concerned about the unintended consequences of AI agents.

When you can’t see how a system makes decisions, anxiety is the natural response. But slowing down AI adoption to mitigate that risk means falling behind the market. The answer to this anxiety is visibility. By bringing agent behavior into your observability practice, you turn unpredictable actions into empirical, governable data.

Cost pressure exposes old habits 

I expected AI infrastructure to dominate our findings on cost pressure. It didn’t come close — only 6% named it their top observability cost driver.

The real culprit is far more familiar. Excessively verbose logs topped the list. We’ve spent years collecting data just in case, sometimes to the point where it breaks budgets. And now that AI gives us a reason to collect still more, the need for an operational data fabric has never been stronger.

There’s a second cost buried in this habit, and it’s the one that concerns me more. When technical exhaust consumes a practice, teams never build the link to business outcomes.

Nearly half of respondents admit telemetry data isn’t a key factor in their organization’s strategic planning.

Strategic planning without operational telemetry is just guessing. If the systems tracking every customer interaction can’t inform the roadmap, something is broken well upstream of any incident.

Leaders maintain control among unpredictable systems 

A group of respondents stood apart. We classify this top-performing cohort as “leaders,” a distinction they earned through three habits practiced together.

First, they put AI agents to work fixing and preventing issues, routinely rather than experimentally. Second, they don’t treat AI as just another service to monitor; they expand the paradigm of observability to govern agent behavior and enforce guardrails around model quality, cost, usage, security, and privacy. Finally, they embed telemetry directly into business planning.

This combination changes their posture in three connected ways.

They’ve shifted attention to where problems now originate. Leaders aren’t dramatically better at watching compute or core networks; most competent teams handle that. Instead, leaders separate themselves at the edge and inside AI decision-making, where customer experience is now most exposed. Nearly every leader (94%) is moderately or very confident in isolating edge and branch network issues, compared to roughly two-thirds of everyone else.

And while the gap for AI-related issues seems more modest (68% of leaders report confidence versus 56% of peers) that delta tells a specific story. Advanced AI users know exactly what they are up against, so their confidence is hard-earned. Less experienced teams are likely overestimating their readiness because they haven’t yet hit the true complexity of agentic behavior.

They practice restraint with data. Leaders align the value of the data with the cost to keep it. By sampling and routing data far more aggressively than their peers, they pay only for high-value signals rather than raw noise. Counterintuitively, that yields more usable visibility. They’re more than twice as likely to say all their observability data is readily accessible and immediately queryable. It also shows up on the ledger: leaders are 2.4 times more likely to say their observability ROI significantly exceeds expectations, and more than half (54%) report at least $5 million in annual business value.

They accept complexity to gain control. This is the finding I keep returning to. Leaders describe monitoring AI as harder than their peers do — 57% say it adds complexity to their jobs, against 38% of all other respondents. They make that trade deliberately, and it pays. Ninety-one percent clearly define ownership and accountability for monitoring AI, versus 59% of everyone else, and 76% say they can accurately control AI costs.

Governance looks like friction from the outside. In practice, it’s what allows these teams to move faster than organizations that improvise oversight one incident at a time. And it changes their teams’ day-to-day: leaders are far more likely to say AI lets them prioritize innovation over routine maintenance, with 37% strongly agreeing versus 22% of their peers.

Observability becomes the north star 

You cannot govern what you cannot see. And the quality of what AI produces depends entirely on the quality of what you feed it.

Observability is how we turn agent behavior into knowledge rather than folklore, how we earn enough trust to hand over real work, and how we keep AI spend tethered to value the business can name. The teams waiting for clarity before they invest have the sequence backward. Clarity is what observability produces.

State of Observability 2026

Get the full research report for complete findings, industry and country breakdowns, and our experts’ recommendations for navigating the agentic future.

More info and download the report.