Observability software company Dynatrace has completed its $915 million acquisition of Arize, a deal first announced in mid-August that aims to merge traditional application monitoring with AI-specific observability. The transaction, which officially closed recently, brings together Dynatrace's decades of experience watching applications and infrastructure with Arize's newer focus on tracing and debugging the behavior of machine learning models and AI agents. The combined platform is meant to solve a growing problem: an application can run flawlessly from a technical standpoint while the AI agent layered on top delivers incorrect answers, invokes the wrong tools, or simply fails to finish its assigned work.
Dynatrace, founded in Austria in 2005 and publicly traded since 2019, has built its business around application performance monitoring and full-stack observability, serving site reliability engineering and platform teams. Arize emerged from stealth in 2020 focused on helping companies monitor machine learning models in production, later expanding into evaluating LLM and agent behavior as those technologies became widespread. The two companies had mutual customers before the deal, and some had already asked to see the platforms function as a unified system, according to Dynatrace CPO Steve Tack. Arize will continue operating both Phoenix, its open-source platform for tracing AI applications, and AX, its enterprise offering, as standalone products even as capabilities migrate into Dynatrace over time.
Arize co-founder and CPO Aparna Dhinakaran says the merger gives developers "both sides of the coin" to investigate failures: the ability to debug LLM-related issues and also trace problems back through the underlying software services that agents depend on. Tack notes that building the assets Arize had already developed, including OpenInference, which extends OpenTelemetry with conventions for capturing LLM calls and agent behavior, "are not easy things to build." IDC group VP Stephen Elliot tells The New Stack that as agents spread across enterprises, aligning these two distinct but related parts of the AI application stack has become increasingly urgent, with combined evaluation and observability helping teams "catch issues earlier and resolve them faster."
The acquisition reflects a broader shift in how observability data gets used. Dhinakaran argues that agents, not humans, will become "the primary consumers" of telemetry, sifting through billions of traces that no engineer would manually inspect. Modern systems generate far more data than people can reasonably review, creating an opening for autonomous agents to continuously analyze that information and act on what they discover. Arize already uses this approach internally with Alyx, an assistant whose traces are reviewed by Signal, another Arize agent that identifies recurring failures and opens pull requests to fix them—around 65% to 70% of which the company accepts, according to Dhinakaran. She describes observability evolving from passively looking at data to actively taking action based on it, with engineers shifting from combing through traces to reviewing code changes proposed by agents.
The long-term vision, as Dhinakaran frames it, is "self-sustaining, self-maintaining software, but also kind of self-improving," with humans retaining final approval. Tack says a conventional partnership might have enabled some integration, but the level of seamless end-to-end experience Dynatrace wanted wouldn't have been achievable without a full acquisition. The ultimate goal is connecting two stages usually treated separately: evaluating AI applications during development and understanding their behavior once deployed in production. The combined platform is designed to trace agent failures across both the AI layer and the conventional software stack beneath it, addressing the reality that autonomous systems don't operate in isolation but depend on the broader ecosystem of applications, services, and infrastructure around them. For enterprises racing to deploy agents at scale, the wager is that debugging intelligence requires visibility into everything from the model's reasoning down to the API it calls—and that separating those views no longer makes sense.