AI Observability Tools Consolidate as Dynatrace Closes Arize Deal

A shortlist that the market has already changed

In mid-July, CRN published its list of the most notable AI observability and governance tools of 2026 so far. Three months later, the list reads as a record of how fast this category is merging into larger platforms. The publication picked ten products: Arize AX, Braintrust, Comet Opik, the Covasant Agent Management Suite, Datadog Agent Observability, Dynatrace AI Observability, JetStream Security's governance platform, LangChain's LangSmith, New Relic AI Observability and Pendo Agent Analytics.1 Two of those names, Arize and Dynatrace, are now one company. Dynatrace said in August that it would buy Arize in a cash-and-stock deal worth $915 million, and it has since closed the purchase.31

That matters well beyond one vendor. The list was meant to celebrate a varied field of specialists. The months that followed suggest those specialists are turning into features inside the platforms that big companies already pay for. Anyone choosing tools for LLM operations should now ask who will own a vendor next year, not only what the product can do today.

What the tools actually do

The products on the list differ in focus, but they share one basic design. They capture structured traces of how an LLM application or agent runs, score those traces with automated evaluations, and track cost and latency in production. Arize AX, for example, records how multi-step agent workflows run and how they call tools, using the OpenTelemetry and OpenInference standards. It then runs evaluations at scale to measure hallucination rates, toxicity and correctness.1 Datadog lists trace-level visibility, automated quality evaluations, behavior clustering, experiments, and cost and performance tracking as the core of its agent product.1 Dynatrace describes distributed tracing, AI guardrails, and cost and token governance, with support for more than 20 AI platforms including OpenAI, Amazon Bedrock, Anthropic and LangChain.1

The main difference is who each tool is built for. Braintrust calls itself the evaluation and observability layer for production AI. In February it added Topics, a feature that automatically groups production traces by recurring patterns, and it raised an $80 million Series B led by ICONIQ in the same month.1 One industry guide puts that round at an $800 million valuation.16 Comet has pushed Opik toward cost management. In June it added spending visibility for coding agents such as Anthropic's Claude Code and OpenAI's Codex.1 New Relic made a similar move with AI Coding Observability, which brings monitoring into the coding phase so teams can measure what AI assistants deliver and what they cost.1 Both launches point to the same trend: coding agents have become a large line item that finance teams want to see.

Evaluation is moving into the release pipeline

The biggest change in practice is that evaluation is no longer a one-off check before launch. It now runs continuously. Vendor guides describe two connected modes. Online scorers grade live production traces. Offline suites run agents against curated test cases, and the same scoring definitions are used locally, in CI and in production.11 One market overview says the standard pattern in 2026 pairs tracing with "eval gates," meaning automated scorers that can stop a regression from shipping.15 Langfuse added experiments with CI/CD quality gates in GitHub Actions in May, and added monitors and alerts in June.2024

Large vendors are following the same approach. IBM's Instana added LLM-as-a-judge evaluations for accuracy and relevance in its AI Agent and LLM Observability release, which is in public preview.18 Snowflake's Observe has announced agent observability, coming soon to private preview, that will run online LLM-as-judge evaluations on production traffic and turn the results into metrics teams can track.19 Using one model to grade another is now standard practice across the industry.

Prompt management is going through the same change. Arize AX includes a prompt and experimentation workspace for testing prompt and agent configurations against production data.1 Langfuse offers prompt version control and label-based deployment, so teams can change prompts without changing code.13 MLflow's guide argues that prompt versioning is now required because small wording changes can quietly lower quality.14 Prompts are starting to be managed like deployable code, with versions, gates and rollback.

Where the sources disagree: depth versus breadth

The coverage is most divided on whether general application monitoring (APM) vendors can match the specialists. Kosmoy's comparison praises Datadog for linking LLM traces with infrastructure and security data and for AI Guard's runtime blocking. It also notes that Datadog offers no self-hosted deployment.20 The same review found no documented evaluation framework or red-teaming in New Relic's AI monitoring as of mid-July.20 CRN's description of New Relic is more generous, citing hallucination and drift detection and enterprise guardrails.1 Latitude takes a middle view: the APM vendors add agent monitoring to existing infrastructure but give up evaluation depth and the ability to find subtle failures.17

The two views can both be true because they measure different things. The APM vendors win on convenience. Kosmoy's own advice is that a company already standardized on Datadog or New Relic is usually better off extending that platform than adding a vendor.24 The specialists win on the evaluation workflow. My view is that this gap is exactly why the acquisitions are happening. Building eval-first tooling inside a company takes a long time, and buying it is quicker.

The consolidation wave

The deals tell that story. ClickHouse bought Langfuse in January, alongside a $400 million Series D that valued ClickHouse at $15 billion. At the time, Langfuse reported more than 2,000 paying customers.26 Cisco announced on April 9 that it would acquire Galileo,16 and plans to combine Galileo's evaluation and guardrail technology with Splunk Observability.1 Lyzr reports that the Galileo deal closed in May,22 while CRN describes it as an April acquisition.1 Datadog bought Adaptive ML on June 30 to speed up its work on agentic post-training.1 Kosmoy counts four ownership changes among tracing vendors in twelve months. Mintlify took Helicone and put it into maintenance mode, CoreWeave absorbed Weights & Biases, and Palo Alto Networks bought Portkey.24

The Arize deal is the clearest example of why. Dynatrace argues that AI engineers judge output quality in one set of tools while SRE teams watch infrastructure in another, so a failed transaction gives no clear path back to the prompt that caused it. Futurum Group analyst Mitch Ashley described the $915 million as the price of that missing link.32 Dynatrace expects the deal to add about 200 basis points to annual recurring revenue growth in fiscal 2027 and to cut non-GAAP operating margin by about 175 basis points. Even the date the deal closed varies by source. Developer Tech reported completion on October 2, and Dynatrace posted about it on October 1. Dynatrace's own blog post went up later.31

Dynatrace says Arize AX stays available as a standalone product and Phoenix remains open source.31 Promises like this now come with almost every deal in the category. ClickHouse made similar public commitments about Langfuse's license and self-hosting.24 Buyers have some protection: Monte Carlo points out that MIT-licensed Langfuse can be forked if its roadmap goes badly.27 The Helicone case shows the other outcome, where an acquired tool keeps working but stops getting new features.24

Standards and governance for agent systems

Open standards are developing alongside the deals. OpenTelemetry formally accepted a code grant of Arize's OpenInference GenAI instrumentation in June, and that code is being merged into the project step by step. OpenTelemetry v1.39 also added Model Context Protocol attributes to tool-execution spans, giving traces visibility into a layer of agent systems that was previously hidden.15 For buyers, shared conventions are the best protection against being stuck with an acquired vendor. If the instrumentation is standard, the backend that stores the traces can be replaced.

The governance products on CRN's list deal with a separate problem: whether an agent should be allowed to act at all. Covasant's suite combines an organization-wide agent registry with a "Control Tower" that audits every LLM call and enforces guardrails.1 JetStream, founded by former CrowdStrike and SentinelOne executives, came out of stealth in March with $34 million in seed funding to sell what it calls a control plane for AI.1 Kosmoy's scoring shows how little the leading observability tools do here. Several get zero for agent containment.20 That score comes from a vendor that sells containment, but the gap it points to looks real.

The reading

The adoption numbers explain the urgency. One market estimate puts LLM observability at about $2.69 billion in 2026 and cites Gartner's estimate that only around 15 percent of GenAI deployments were instrumented in early 2026.15 In a Dynatrace survey of agentic AI leaders, 51 percent named managing and monitoring agents at scale as a top barrier to putting agents into production.35

The CRN list is best read as a picture of a category that was splitting apart while it grew. Tracing has become a commodity, standardized through OpenTelemetry and available free in self-hosted open-source tools. The value is moving to two places: evaluation that sits inside the release pipeline, and policy control over agents that take real actions. Large platform vendors are buying their way into the first, and startups are competing for the second. Teams running LLM operations in production should build around open instrumentation, treat evaluation suites and prompt versions as their own assets, and plan for the vendor behind their dashboard to change hands.