From dashboards to decisions: Why AI agents can’t function without observability
For most of its history, observability has been a human-facing discipline: dashboards with red, yellow and green lights, engineers staring at graphs, alerts routed to a war room. That era, according to Philippe Deblois, Global Vice President of Solutions Engineering at Dynatrace, is ending faster than most enterprises realize.
“AI is the big topic of the day, and it introduces a whole lot more complexity to pretty much any project, any workload that gets submitted into organisations,” Deblois said. “Without visibility, it’s very hard to manage how to automate the actions we want to take based on AI. So observability is becoming more and more critical.”
That single shift — from observability as a human convenience to observability as the precondition for AI to act at all — is reshaping how the discipline is built, sold and used across enterprises. It also explains why, in Deblois’s telling, the last two years look nothing like the two years ahead.
From AI-assisted to AI-built: a two-team problem
Until recently, Deblois said, the industry’s use of AI within observability platforms was mostly internal plumbing. Root-cause analysis, anomaly detection, correlation — AI quietly made these faster and more accurate, but the workload being observed was still conventional software.
“We would say that observability was AI-assisted in the past couple of years,” he said. “What’s changed now is that people are building with AI, and as such they need observability to observe AI itself. So you had observability that was powered by AI, and now you have observability that’s built for AI.”
That distinction is more than semantic. It has created what Deblois described as a structural split inside enterprises: AI teams building and observing their own models and pipelines, and central IT/SRE teams using AI-powered observability to resolve incidents faster — mean-time-to-resolve, mean-time-to-detect, faster root-cause analysis — but not necessarily observing AI systems themselves. Two teams, two use cases, converging on the same telemetry.
The agentic shift: visibility becomes existential
If AI-built systems complicated observability, autonomous agents have upended it. When agents call other agents, which call still more agents, the resulting chains of decisions are opaque by default — and, as Deblois pointed out, increasingly outside any single team’s direct control.
“If an agent is talking to another agent, to another agent, you need that visibility to understand, almost in a replay, what exactly happened,” he said. But the more consequential change, he argued, is that agents themselves have become consumers of observability data, not just its subject. “Agents now need to leverage observability to do their thinking. An agent might decide it needs to reason about what the next outcome should be, and it’s going to leverage the telemetry collected by an observability solution to make decisions that then lead to automation and action.”
In other words, telemetry is no longer just evidence for humans after the fact. It is becoming the raw material an autonomous system uses to decide what to do next.
Human-in-the-loop: a bridge, not a destination
That raises an obvious governance question: who — or what — keeps agents inside their guardrails? Deblois’s answer is candid about where the industry actually stands, as opposed to where the marketing often suggests it is.
“Not many customers that I meet across the world are fully autonomous today,” he said. “We call it supervised autonomous operations, where a human is still there to click the button and say, ‘yes, we approve.’ I think that’ll be there for a while, until people have confidence that the AI can handle it on its own.”
Dynatrace, he said, has built its platform to support full autonomy technically — but adoption is deliberately paced by trust, not capability. “Most customers want to have those gates today, to build that confidence that the AI is going to do what it promises to do. All those gates can be automated within the platform, so a human presses the button and says, ‘yes, this makes sense, move on to the next step.'”
Pressed on whether human-in-the-loop can survive the sheer velocity at which agentic systems now operate, Deblois acknowledged the tension directly: it is, he said, “a very delicate balance between the power of AI and the speed we want to operate at.” His expectation is that the balance will shift gradually — from human approval on every action today, toward selective, risk-weighted autonomy as trust is earned incident by incident. What won’t change, he argued, is the dependency underneath it: agents will still need enough context, pulled from observability platforms, to make decisions in a way that’s trustworthy in the first place.
The deterministic core inside probabilistic decisions
That trust question is where Deblois drew the sharpest architectural line of the conversation — and arguably Dynatrace’s clearest point of differentiation. Large language models reason probabilistically. Infrastructure, by contrast, cannot afford to be probabilistic about what actually happened.
“From our perspective at Dynatrace, we use deterministic AI to be the foundation for these agents to leverage from an observability standpoint. That gives more precise answers,” he said, pointing to
Dynatrace’s dependency-mapping engine, SmartScape, which maintains a real-time graph of dependencies across an entire call chain. “If you think of a user clicking a button that’s a prompt to another agent, and that agent talks to another agent, and talks to another agent — we can see that entire call end to end. When there’s a problem, we follow that trace across all those systems and understand the possible causes. That’s the deterministic aspect. We actually see the transaction, across millions of transactions, end to end.”
The implication is a kind of division of labor: agents reason probabilistically about what to do, but they ground that reasoning in a deterministic, verifiable record of what the system actually did. Without that grounding, Deblois suggested, agentic AI is reasoning on top of guesswork rather than fact.
Learning from agents that went rogue
The conversation turned to a question that has become uncomfortably familiar to AI leaders this year: publicized incidents in which AI agents at major labs bypassed sandbox restrictions and exceeded their intended limits. Deblois was careful not to overclaim — “I’m not saying your platform prevents that right now, because it’s still a very evolving phase” — but he laid out what he sees as the necessary architecture to prevent recurrence.
“Visibility is the key. You need a visibility layer first — observability becomes that layer to understand what agents are doing,” he said. “Then you have to have guardrails in place: if an agent does X, Y, Z, that’s out of bounds. Once you have visibility and guardrails in place, you can automate the prevention of going rogue. If something goes off those guardrails, stop it from moving forward.”
His formulation — visibility, guardrails, automated enforcement, in that order — is a useful checklist for CIOs and CISOs currently drafting agentic AI governance policies, many of whom are discovering that traditional access-control models weren’t designed for software that makes autonomous, multi-step decisions.
Observability moves into the developer’s workflow — and beyond the dashboard
AI’s influence on observability isn’t confined to agent governance. It’s also changing where and how humans interact with telemetry at all. Deblois pointed to the rise of AI coding assistants as a case in point: developers using tools like Claude Code now need real-time observability data available inside their existing workflow, not in a separate console.
“Developers need access to real-time observability data to debug the code they may have produced through an AI coding agent,” he said. “They never even have to go into a data trace or an observability platform. They can stay right within their coding tool and say, ‘I’m debugging this service — please provide me the telemetry to help me solve this problem.'”
That points to a broader trend Deblois expects to accelerate sharply: the disappearance of the dashboard itself. “The world of observability today is very human-led still — people watching dashboards, red, green, yellow lights,” he said. “I see a world, even as early as next year, where people are leveraging observability data in a headless way — never needing a dashboard, just pulling the insights directly from whatever flow they’re in.” An incident manager working in ITSM, he said, would pull telemetry via MCP or CLI without ever opening a Dynatrace console; a developer would do the same from inside an IDE.
MCP’s coming-of-age moment
That headless future depends heavily on the Model Context Protocol (MCP), the emerging standard for connecting AI models and agents to external tools and data sources. Deblois sees MCP adoption accelerating quickly — alongside growing use of command-line interfaces by enterprise customers — but argues the protocol itself still has real maturing to do, particularly around enterprise security.
“Enterprise customers will leverage MCP primarily for reasons like security — making sure there are guardrails around what it can and cannot communicate with,” he said. “As an industry, all vendors are now exposing things through MCP. A vendor like Dynatrace might need to talk to ServiceNow and get context from it — that’s increasingly going to happen through MCP.” The next step, in his view, is hardening the protocol itself: “adding more enterprise features to make sure no APIs leak, none of those security guardrails are breached.”
It’s a telling parallel to draw, and one Deblois invoked explicitly in the conversation: MCP today occupies roughly the position APIs did a couple of decades ago — a powerful but immature integration layer that the industry will spend the next several years hardening into something enterprises can trust by default.
Full-loop SRE automation is closer than it looks
Asked to look one or two years out rather than five — a timeframe he considers more honest, given the pace of change — Deblois pointed to full-loop site reliability engineering automation as one of the most concrete near-term shifts. Rather than AI merely accelerating a human SRE’s work, he expects AI systems to close the loop end to end: finding root cause, communicating with other systems and tools — including hyperscaler AI SREs — determining the actual problem, and automatically generating the pull request to fix it.
“You can see that as definitely something within the next year or two, where most customers will start to leverage that in their environments,” he said.
Beyond IT: observability goes into ovens, cars and factory floors
Perhaps the most striking part of the conversation was how far Deblois expects observability to travel beyond its traditional IT home. With the maturation of OpenTelemetry as an open instrumentation standard, he said, virtually any system with a sensor can now emit usable telemetry — and enterprises are already acting on it.
“We have customers now leveraging telemetry in cars. I had a discussion with a restaurant business — an oven can emit telemetry now,” he said. “Being able to capture that at a micro level becomes very important, because if an oven doesn’t work, you’re not serving your burgers to the customer. It’s an immediate impact to business.”
That expansion into operational technology (OT) — manufacturing plants, healthcare equipment, industrial control systems, retail devices — represents, in Deblois’s view, one of the largest untapped growth areas for the category. “The opportunity is huge for the observability space because it’s expanding into new use cases, new verticals — and it’s adding the business dimension to things.”
That “business dimension” is its own emerging use case: real-time business observability. Traditionally, business metrics have lagged operational reality by hours or days, arriving only after data has been collected and crunched. Deblois argues that when telemetry is pulled directly from the systems generating it, that lag disappears. “I can tell you an oven was down because I see it wasn’t available, and that caused this impact on revenue — and I know that in real time,” he said. “Business observability is another big use case that’s available anywhere IoT is involved.”
Betting on open standards at scale
Dynatrace’s clearest strategic move in this direction is its acquisition of Bindplane, a company specializing in collecting, ingesting and scaling OpenTelemetry data. Deblois framed the deal as a bet on open standards being the only viable path to observability at the scale the next decade will demand.
“A big reason for making that acquisition is that they were a leader in collecting, ingesting and scaling the collection of open telemetry data. It’s an open-source-focused approach that we believe can scale to some of the largest environments,” he said. “We’ve had customers with millions of different endpoints all feeding into this solution — and we have the scalable backend to bring all that data together, because that’s the world we’re coming to.”
The real endgame is an analytics problem and not a collection problem. If there’s a single thesis running through Deblois’s view of where observability is headed, it’s that collecting telemetry was never really the hard part — making sense of it is. As data volume explodes across IT, AI agents, and now OT and IoT devices, the bottleneck shifts decisively from ingestion to interpretation.
“At the end of the day, a lot of customers are going to be solving analytics problems more than they will be just collecting telemetry,” he said. “That’s one thing — to collect it. But how do you make sense of it, and how do you understand how to best use it? It becomes an analytics problem.”
The takeaway for enterprise leaders
For CIOs and CTOs building AI strategies, Deblois’s remarks point to a few clear priorities: treat observability as infrastructure for AI decision-making, not just IT monitoring; invest in deterministic, verifiable telemetry as the grounding layer beneath probabilistic agent reasoning; harden MCP and other agent-integration standards before scaling autonomous operations; and start planning for telemetry that extends well beyond servers and applications into physical operations. The organisations that get ahead on these fronts, in his view, won’t just detect problems faster — they’ll be the ones whose AI agents can actually be trusted to act on their own.