Dynatrace will acquire AI observability and agent development startup Arize AI for $915 million in cash and stock, the company said. The deal is Dynatrace's third acquisition this year and signals its bet that enterprise demand for custom AI agent development will grow quickly enough to justify a nine-figure purchase price.
Arize, a 200-person company based in Berkeley, Calif., was founded in 2020. Its AX platform evaluates and experiments on AI agents, using trace data to monitor and refine large language model output quality. Dynatrace plans to combine Arize with Bluebox AI, its recently launched product that automates the software development lifecycle for regular applications.
Together, the tools form an automated improvement cycle for AI agents: Dynatrace supplies runtime telemetry, Arize evaluates model outputs, and Bluebox fixes the underlying agents based on those findings, said Torsten Volk, an analyst at Omdia.
A bet on agentic software development
Enterprises are increasingly looking to build and monitor AI agents alongside traditional software, and Dynatrace is positioning itself as the platform for both. The company described the strategy as moving from observability toward "active systems of control," a direction it signaled with its January acquisition of feature flagging startup DevCycle and its April purchase of data pipeline vendor Bindplane.
"Arize was built as an agent-first platform, with one of the largest collections of agent trace data in the industry," said Arize co-founder and CEO Jason Lopatecki in a blog post. "Dynatrace brings depth in tracing and logging software systems that provide the data to drive these agents. These two areas just belong together if we are going to build the future systems that continuously improve."
Dynatrace's earlier AI observability "was just focused on standard metrics, not on actually looking at model outputs in different constellations," Volk said.
One early customer reaction reflects the perceived gap. "DIY agents are now like websites: We're gonna have millions of them, and most of them are untested and will not scale," said Mark Tomlinson, AVP of performance and observability at FreedomPay, in an email. "Bluebox + Arize is a nice combination that bolsters Dynatrace's long-standing commitment to 'find and fix' value delivery."
Vendors flock to agent evaluation
Dynatrace is more of an outlier here. Cisco acquired AI agent evaluation startup Galileo in April, integrating it with Splunk's observability tools. Harness added AI agent development support to its DevSecOps pipelines in July. Vendors are responding to survey data: an ALM & Observability survey of 400 IT professionals last year found 49% of respondents ranked monitoring AI and ML systems with the same platform used for infrastructure and applications among their top priorities.
"Agent debugging, evaluations, agent root cause analysis and testing - might even be more important than they were in the traditional SDLC model, especially with agentic systems that make decisions," said Stephen Elliot, analyst at IDC. "It is a foundational investment for technology executives."
For IT and development pros, the trend points to a shift in daily work. Agent development carries its own failure modes, including model drift and undocumented behavior, and debugging them requires the same rigor that AD&D teams apply to traditional code. The practical question being worked out is whether the tools for that discipline come from an observability vendor, a DevOps pipeline vendor, or a dedicated agent AI vendor.
Those who need to start building these skills can explore the AI for Software Developers training path, while AI for IT Managers covers the strategic side of integrating agentic systems into existing IT operations.
Why this matters for IT and Development
Enterprises at scale will soon run thousands of AI agents, and each structured decision or generated output needs to be traced, evaluated, and corrected if it is to be trustworthy in production. This acquisition signals that IT teams will be expected to manage AI agents as part of normal SDLC work - with dedicated evaluation stages and automated QA - rather than treating them as a separate, experimental effort.
Your membership also unlocks:
