In this article, “AI agents” refers to customer-deployed agents that reason over context, select tools, take actions, or coordinate with other agents.
In this article, “AI agents” refers to customer-deployed agents that reason over context, select tools, take actions, or coordinate with other agents.
Enterprises are increasingly moving AI agents from experimentation into production. But an agent can be available and still produce an inaccurate answer or can complete a workflow while taking the wrong path. Conventional monitoring alone may not reveal whether an agent used appropriate context, tools or policies. This post draws on the new whitepaper from Deloitte and Dynatrace and connects its findings to what teams are experiencing in practice.
Agentic AI observability connects an agent’s context, decisions, actions, and outcomes with the applications and infrastructure it depends on. That evidence helps teams determine whether an agent is working as intended, not merely whether it is running. Only 21% of organizations have a mature governance model for autonomous agents, according to Deloitte’s State of AI in Enterprise report.
Key takeaways
- Design for the right level of autonomy. Increase it only when performance and control thresholds are consistently met, with policy gates and human oversight.
- Build in observability by design. Define traceability, quality thresholds, cost controls, and escalation paths before agents become critical.
- Standardize evidence across AI and production. Common telemetry, metadata, and decision lineage help connect AI behavior with business outcomes.
- Evolve the operating model as autonomy scales. Redefine decision rights, oversight, and accountability across business, engineering, security, and risk.
Conventional monitoring falls short for AI agents
Customer deployed AI agents are not like conventional software: they can reason over context, select tools, and interact with other agents. The same task can follow different execution paths and produce different outcomes. Conventional service and API monitoring shows whether a service has responded, or an API failed. It can’t show whether an agent used the right context or tool, followed policy, or produced an outcome worth the cost.
The whitepaper calls this the agentic observability gap. The goal isn’t another layer of telemetry. It’s to connect the context an agent retrieved, the decision it made, the action it took, and the business impact.
AI adoption is growing faster than visibiliby and control
The paper’s findings show up in field data. In the Dynatrace Pulse of Agentic AI Report 2026 which surveyed 919 IT and AI executives, 51% named managing and monitoring AI agents at scale as their top operational challenge, and 45% said they lack clear rules for when agents can act autonomously. Also, 42% have limited real-time visibility to trace and troubleshoot agent behavior.
AI engineering and production operations evolved as separate disciplines with separate tools. AI engineers see agent behavior through traces, evaluations, and experiments. Platform and SRE teams monitor the applications, cloud complexity, and infrastructure beneath. Neither side can fully explain how AI behavior and system behavior interact, and the cause of a failure often sits outside the model: a slow upstream service, a degraded retrieval result, or an infrastructure constraint.
Match observability to the level of autonomy
For customer-deployed agents, the whitepaper describes four levels: At suggestion, AI advises and a human decides, so the focus is usage, cost, and output quality. At embedded, AI output enters the process and can degrade silently, so teams need drift detection and business outcome correlation. At action, agents execute decisions, so decision traceability and guardrails matter. At systemic scale, where multiple agents may coordinate, teams may need cross-agent correlation and supporting causal evidence. Understand the impact of each step before you define the next observability baseline.
To help teams assess where they stand, the whitepaper also introduces the Enterprise Agentic Scale framework. It presents an assessment framework organized around six dimensions, 12 subdimensions, and 60 capabilities, from observability strategy and AI telemetry coverage to metadata standardization, AI-powered correlation, risk and governance, and operational integration.
Connect agent evidence with production context
Operating AI with confidence means evaluating it, observing it in production, and acting on what you learn. Dynatrace positions its capabilities around: Built for AI and Powered by AI.
- Evaluate with Arize. Arize goes deeper on AI engineering workflows, including evaluation, agent investigation, and experimentation, through Arize AX and the open-source Phoenix.
- Observe with Dynatrace. Dynatrace provides broader production context across applications and infrastructure, including performance, cost, quality, and guardrail adherence.
- Act with Dynatrace Intelligence. Dynatrace Intelligence combines deterministic, causal AI with agentic AI to deliver precise answers. It powers automated root cause analysis, anomaly detection, and autonomous agentic operations.
These roles complement each other, and customers can use Arize, Dynatrace, or both. Arize helps answer whether the agent is accurate and improving. Dynatrace helps answer whether the system around it is running reliably, and what it costs.
Connecting agent evidence with production context gives teams a common basis for investigating AI quality, system reliability, performance, and cost. Dynatrace and Arize offer complementary capabilities while product integration continues following the October 1, 2026, acquisition.
Four outcomes teams can work toward
Dynatrace and Arize help enterprises build, run, and continuously improve AI they can stand behind.
- Cost and performance: Correlate token consumption with execution context to find inefficient reasoning loops and high-cost prompts.
- Quality and accuracy: Understand agent behavior, evaluate quality, investigate failures, and test changes.
- Governance and compliance: Monitor the guardrails your app and model providers enforce and preserve lineage from context to decision to outcome to support audit evidence for regulations such as the EU AI Act.
- Supervised autonomous operations: With Dynatrace Intelligence, identify root causes across applications, infrastructure, and agents, trigger remediation under human oversight, and escalate exceptions.
Humans set goals, policies, and guardrails. Agents execute within them. Observability provides continuous evidence that connects the two.
“Autonomy should be earned through demonstrated performance and compliance through observability, not assumed because the technology allows it.” – Enterprise Agentic Observability whitepaper, Deloitte
“Autonomy should be earned through demonstrated performance and compliance through observability, not assumed because the technology allows it.”
– Enterprise Agentic Observability whitepaper, Deloitte
To get started, Deloitte and Dynatrace recommend four phases: assess, design, implement, and scale. Prove value in a pilot, then expand autonomy as confidence and control mature.
- Download the whitepaper, Enterprise Agentic Observability.
- Read the announcement:Dynatrace completes acquisition of Arize.












