Datadog AI Agent Monitoring Faces Dynatrace's $915 Million Arize Bet

Datadog AI agent monitoring now faces a sharper challenge after Dynatrace agreed to acquire Arize in a transaction valued at $915 million. The deal turns an emerging product category into a direct contest between two established observability vendors.

Both companies want to become the system enterprises use to inspect AI agents, evaluate their output, control their costs, and investigate failures. However, they are approaching that opportunity from different positions. Datadog is extending its cloud monitoring platform into agent development, while Dynatrace is buying deeper evaluation expertise.

The timing matters because agents create operational problems that ordinary application dashboards were not designed to explain. An agent can select tools, call other agents, retrieve external data, and repeat steps before producing an answer. A successful request can still deliver the wrong business outcome, expose sensitive information, or consume far more resources than expected.

The result is a useful reversal. AI agents promise to automate digital work, yet enterprises need another layer of software to supervise that automation. Datadog and Dynatrace are racing to own that control layer before buyers settle on smaller AI-native specialists.

Dynatrace Turned AI Observability Into a $915 Million Contest

Dynatrace's planned Arize acquisition moves the competition beyond ordinary infrastructure monitoring and into the full AI development lifecycle.

Dynatrace announced the definitive agreement on August 13, 2026. The company said it would acquire Arize for $915 million, subject to customary adjustments. The consideration includes approximately $815 million in cash and replacement equity awards for Arize employees.

The transaction is expected to close after regulatory review and other customary conditions. Until that happens, Dynatrace and Arize remain separate businesses. Their promised integration should therefore be treated as a roadmap, not a completed product result.

Arize specializes in evaluating and observing machine learning systems, large language model applications, and agents. Its tools help developers inspect traces, compare experiments, assess response quality, and identify changes in model behavior.

Dynatrace brings a different foundation. Its platform already connects application traces, infrastructure health, logs, user activity, and business transactions. The company argues that combining those operational signals with Arize's evaluation tools will cover AI applications from development through production.

That combination addresses a real division inside many engineering organizations. AI teams often test prompts and models in one environment. Operations teams monitor latency, errors, infrastructure, and incidents somewhere else.

Dynatrace says the proposed acquisition will close that gap. Its Arize agreement describes AI observability as a lifecycle spanning experimentation, evaluation, production tracing, infrastructure, and business outcomes.

The announcement also reveals how valuable the category has become to established monitoring vendors. Dynatrace is not merely adding another dashboard. It is committing substantial capital to connect AI evaluation with its existing enterprise platform.

The company already launched its dedicated AI Observability app in January 2026. That product supports agent interactions, tool usage, dependencies, token consumption, latency, cost trends, and guardrail outcomes.

Dynatrace also says it supports more than 40 LLM technologies. Its listed integrations include OpenAI, Anthropic, Amazon Bedrock, Google Gemini, LangChain, and emerging agent protocols.

Arize gives Dynatrace a stronger route into AI engineering teams that work before an application reaches production. It also gives the company open-source reach through Arize's Phoenix project and a more specialized evaluation workflow.

Datadog now faces a competitor that is combining acquisition speed with an established enterprise sales channel. Dynatrace faces the harder task of integrating two products without weakening the developer experience that made Arize attractive.

Why AI Agents Break Traditional Monitoring

An agent can remain technically available while making a sequence of individually valid decisions that produces a bad result.

Traditional application monitoring usually asks whether a service responded, how long it took, and whether an error occurred. Those measures remain necessary, but they do not explain an agent's reasoning path or output quality.

Consider a customer support agent processing a refund. It might retrieve an account, consult a policy, call a billing tool, and update a customer record. Every service can return a successful response while the agent applies the wrong policy.

A second run of the same request can follow another path. Large language models are nondeterministic, meaning identical inputs do not always produce identical outputs. That variability weakens tests built around fixed expected results.

Tool use adds another source of uncertainty. An agent can select the wrong function, supply unsuitable arguments, or call the correct tool at the wrong time. It can also enter a retry loop that increases latency and token consumption without generating a conventional application error.

Multi-agent systems expand the problem further. A coordinating agent can delegate work to specialized agents, which then call external models and services. The final answer may depend on several prompts, permissions, retrieval steps, and model responses.

Teams consequently need traces that preserve an entire execution path. A trace is a connected record of the steps taken during one request, including model calls, tool calls, timing, errors, and relevant metadata.

Yet collecting traces is only the beginning. Engineers must connect technical behavior to quality, safety, and business outcomes. A fast response is not useful when it contains an unsupported claim or triggers an unauthorized action.

Datadog described this problem when it launched AI Agent Monitoring in June 2025. Its monitoring release said the product maps inputs, tool invocations, calls to other agents, and outputs in an interactive graph.

The company says engineers can investigate latency spikes, incorrect tool calls, and infinite loops. They can then correlate those events with quality, security, and cost measures.

Dynatrace takes a similarly broad view. Its AI Observability app follows prompts, tool calls, and model invocations while linking them to infrastructure and application dependencies.

These capabilities explain why AI observability is becoming more than an LLM dashboard. The operational object has changed from a predictable software transaction into a probabilistic workflow with authority to act.

That change pressures developers, security teams, site reliability engineers, and business owners simultaneously. Each group needs a different answer from the same evidence.

Developers want to know which prompt or tool caused a failure. Operations teams need latency, reliability, and dependency data. Security teams need permissions and audit trails. Business leaders want evidence that the agent produces useful outcomes.

A platform that connects those questions can become difficult to replace. That is the strategic prize behind the Datadog and Dynatrace contest.

Datadog AI Agent Monitoring Starts With Production Data

Datadog's advantage is its ability to place agent behavior beside the application and infrastructure signals customers already collect.

Datadog introduced AI Agent Monitoring, LLM Experiments, and an AI Agents Console as parts of its broader LLM Observability offering. Together, they cover runtime investigation, controlled testing, and centralized oversight.

AI Agent Monitoring focuses on what happens during an execution. Engineers can examine the chain of model calls, tools, and agent handoffs. The interface also connects those steps with latency, token usage, errors, cost, and evaluation results.

LLM Experiments addresses changes made before deployment. Teams can compare prompts, models, and application configurations using datasets created from production traces or supplied examples.

That connection between production traces and experiments is central to Datadog's approach. A failure discovered in a live workflow can become a repeatable test case. Engineers can then compare a proposed fix against the original behavior.

Datadog's September 2026 guidance illustrates the workflow with a support agent. A team might find truncated summaries and slow billing calls after deployment. Trace data can identify recurring patterns, while experiments test whether a prompt or model change improves them.

The company also recommends comparing nondeterministic outputs across repeated runs. A single successful response provides weak evidence when another execution can produce a different result.

Evaluators add another layer. These are tests that score an agent's output for qualities such as relevance, accuracy, safety, or task completion. Some use deterministic rules, while others use another language model as a judge.

Model-based evaluation can scale review, but it introduces its own uncertainty. Datadog advises calibrating an LLM judge against human assessment before relying on its scores. That caution matters because two models can share blind spots.

The AI Agents Console extends monitoring beyond software built by an enterprise. Datadog says it can help organizations catalog internal and third-party agents, inspect usage, measure impact, and check permissions.

That catalog function becomes important when employees adopt agents embedded in coding tools, customer platforms, and productivity software. An organization may not own the agent's model or runtime, yet it remains responsible for data access and outcomes.

Datadog's broader commercial position gives this strategy weight. The company reported second-quarter 2026 revenue of $1.12 billion, up 36 percent from the prior year. It also reported about 4,720 customers with annual recurring revenue above $100,000.

Those numbers do not prove that Datadog will lead AI observability. They show that it has a large customer base where agent monitoring can become another attached product.

Datadog also made Bits Code, Bits Chat, and Bits Agent Builder generally available during the quarter. These products place the company on both sides of the market. It builds agents for customers while selling the systems used to observe agents.

That position creates useful product feedback. Datadog can apply its monitoring tools to its own agents and encounter operational problems directly. It also creates a credibility test because customers can judge whether the company's autonomous products remain inspectable.

For engineering teams, the appeal is consolidation. An agent incident rarely stops at the model boundary. The root cause might sit in a slow database, a failing API, a retrieval pipeline, or an overloaded GPU.

Datadog can correlate those layers inside an existing monitoring environment. Teams also need searchable technical context outside live telemetry, including incident records and design decisions. A maintained engineering knowledge base can preserve that human context.

Datadog's challenge is proving that a broad platform can match specialist tools on evaluation depth and developer usability. Existing distribution opens the door, but it does not settle the product decision.

Dynatrace Is Buying the Development Side of the Lifecycle

Dynatrace's Arize strategy attacks Datadog by connecting pre-release evaluation with a production platform built around dependency context.

Dynatrace has historically emphasized causal analysis. Its Smartscape technology maps relationships among applications, infrastructure, services, networks, and users. Its Grail data layer consolidates observability, security, and business information.

The company argues that this context helps teams understand why an incident occurred, not merely which metric changed. That approach fits agent systems because a bad result can emerge across several dependencies.

Dynatrace's dedicated AI Observability app already tracks agent execution paths, tool invocations, model interactions, token use, cost, latency, and errors. It also supports OpenTelemetry and OpenLLMetry, which provide open conventions for capturing telemetry.

Open instrumentation matters because the agent stack is changing quickly. Enterprises do not want every new model, framework, or protocol to require a proprietary monitoring integration.

Dynatrace lists support for Amazon Bedrock AgentCore, Strands Agents, LangChain, Google's Agent Development Kit, the OpenAI Agents SDK, and Model Context Protocol workflows. This breadth reduces friction for companies running mixed environments.

Arize fills a different need. AI engineers use evaluation systems while choosing datasets, prompts, models, and release candidates. Their workflows begin before operations teams receive a production service.

The proposed combination therefore targets a persistent organizational split. Development teams investigate model behavior and response quality. Operations teams examine availability, application dependencies, and infrastructure performance.

Bringing those activities together could create a continuous improvement loop. A production failure becomes an evaluation case. The resulting change can be tested before a controlled deployment returns it to production.

Dynatrace says the combined platform will connect AI behavior with business impact. That is an important claim because technical proxies can hide failure.

For example, an agent might reduce average handling time while increasing reopened support cases. A coding agent might produce more changes while increasing rollback rates. An observability platform needs outcome data to distinguish activity from value.

Dynatrace entered this contest with meaningful momentum. It reported $2.136 billion in annual recurring revenue for its first fiscal quarter of 2027, an increase of 17 percent.

The company also reported quarterly revenue of $555 million, up 16 percent. Organic net new annual recurring revenue grew 41 percent, according to its quarterly results.

Dynatrace says demand is rising as customers expand cloud-native workloads and AI initiatives. However, company commentary cannot reveal how much revenue comes directly from AI observability.

The Arize transaction is therefore both a product decision and a market wager. Dynatrace is paying for technology, talent, developer adoption, and time.

The integration risk is substantial. Enterprise monitoring platforms and developer-focused evaluation tools serve different users. Their interfaces, deployment patterns, release cycles, and buying processes do not automatically align.

Dynatrace must preserve Arize's appeal to AI engineers while connecting it with the larger platform. A forced migration or overly complex packaging could push developers toward independent alternatives.

It must also show that the combined data produces better decisions. More telemetry does not automatically create more understanding. Poor schemas, incomplete traces, and inconsistent evaluation criteria can produce a larger collection of ambiguous evidence.

The Winner Must Monitor Quality, Cost, and Authority

The decisive feature will not be the prettiest trace graph, but the ability to connect agent actions with acceptable business outcomes.

Datadog and Dynatrace can both display execution paths, token use, errors, and latency. Those capabilities are becoming baseline requirements rather than lasting differentiators.

The harder problem is defining success. An agent can complete a workflow without satisfying the user. It can deliver an accurate answer while violating a policy. It can improve quality while making the workflow uneconomical.

Quality evaluation remains especially difficult. Rules can verify structured outputs, required fields, and known facts. Open-ended writing, reasoning, and recommendations require more subjective judgments.

LLM-as-a-judge evaluation offers scale, but it should not become unquestioned ground truth. A judge model can prefer certain writing styles, miss subtle factual errors, or reproduce biases found in the evaluated model.

Human review remains necessary for calibration and high-risk cases. However, reviewing every execution defeats much of the economic case for automation. Vendors must help customers decide which cases deserve human attention.

Cost is easier to count but harder to interpret. Token consumption, model fees, infrastructure use, and external tool calls can all contribute to an agent's operating cost.

A retry loop can inflate spending without producing an obvious outage. A more expensive model can still be economical when it completes tasks reliably and reduces rework. Cost monitoring therefore needs an outcome denominator.

Authority creates the highest stakes. Agents can access customer records, repositories, financial systems, and communication tools. Observability must capture what an agent attempted, which permissions it used, and whether an action was approved.

Recording an unsafe action after completion is not enough for sensitive workflows. Enterprises will increasingly demand policy enforcement, revocable access, human approval, and recoverable execution.

This requirement blurs the boundary between observability, security, governance, and runtime control. Datadog and Dynatrace are not only competing with each other. They face specialists focused on evaluations, agent gateways, security, tracing, and workflow orchestration.

OpenTelemetry can reduce vendor lock-in at the collection layer. It cannot standardize every evaluation score or business outcome. Vendors will still make different choices about data models, correlations, and automated recommendations.

That leaves buyers with a skeptical question: does a unified platform reveal more, or merely centralize more data?

The answer will vary by organization. A company already standardized on Datadog may value rapid deployment and familiar workflows. A large hybrid enterprise may prefer Dynatrace's dependency model and causal analysis.

AI-native teams may continue choosing specialist evaluation tools. They can send selected telemetry into a broader monitoring platform later. Dynatrace's Arize purchase is an attempt to prevent that specialist layer from remaining separate.

Neither vendor has established a universal method for proving agent reliability. Their products can collect evidence, run tests, and highlight anomalies. Customers must still define acceptable behavior and meaningful outcomes.

The risk is that AI observability becomes a collection of attractive dashboards without operational accountability. Successful deployments will link every important score to a decision, owner, threshold, and response.

Three Signals Will Decide the Datadog and Dynatrace Race

The next phase will be judged by integration, measurable adoption, and the degree of control enterprises grant these platforms.

The first signal is the completion and product integration of Dynatrace's Arize acquisition. Closing the deal only starts the test. Buyers should watch how quickly traces, datasets, evaluations, and production incidents move between the two environments.

A coherent workflow would strengthen Dynatrace's lifecycle thesis. Separate interfaces, duplicated data, or unclear product ownership would weaken it. Developer retention around Arize and Phoenix will provide another useful signal.

The second signal is customer adoption disclosed by both vendors. General revenue growth offers context, but it does not isolate demand for agent observability.

Useful evidence would include the number of production agents monitored, workloads using evaluations, attached products per customer, and customer expansion tied to AI. Case studies should identify measured outcomes rather than agent counts alone.

The third signal is movement from observation to governed action. Both companies are building agents that can investigate incidents and assist operations. Customers must decide whether those systems can recommend changes, execute approved steps, or act autonomously.

Read-only analysis carries a lower trust burden. Automated remediation requires complete audit trails, constrained permissions, rollback mechanisms, and clear responsibility when the agent makes a mistake.

Progress on governed action would strengthen the argument that observability platforms can become control planes for enterprise agents. Persistent hesitation would suggest that monitoring remains valuable but autonomy has reached a trust boundary.

Enterprise buyers should not choose a platform from a feature checklist alone. They should begin with one production workflow and define its acceptable quality, cost, latency, permissions, and business result.

They should then test whether the platform can reconstruct a failure, identify its cause, compare a proposed fix, and prevent regression. That exercise exposes gaps that a polished demonstration can conceal.

Datadog AI agent monitoring currently offers a strong production-first route for organizations already using its platform. Dynatrace is making a larger bet that evaluation and operations must become one lifecycle.

The winner will be the vendor that turns unpredictable agent behavior into evidence teams can trust and act upon. For buyers, the immediate question is simpler: can your monitoring system explain not only whether an agent ran, but whether it did the right thing?