AzureFeeds

If you’ve shipped an AI agent to production, you’ve likely run into the same uncomfortable realization: the hard part isn’t getting the agent to work – it’s keeping it working. Models get updated, prompts get tweaked, retrieval pipelines drift, and user traffic surfaces edge cases that never appeared in your eval suite. Quality isn’t something you establish once. It’s something you have to continuously measure.

Today, we’re making that continuous measurement a first-class operational capability. Evaluations, Monitoring, and Tracing in Microsoft Foundry are now generally available through Foundry Control Plane. These aren’t standalone tools bolted onto the side of the platform – they’re deeply integrated with Azure Monitor, which means AI agent observability now lives in the same operational plane as the rest of your infrastructure.

The Problem With Point-in-Time Evaluation

Most evaluation workflows are designed around a pre-deployment gate. You build a test dataset, run your evals, review the scores, and ship. That approach has real value – but it has a hard ceiling.

In production, agent behavior is a function of many things that change independently of your code:

• Foundation model updates ship continuously and can shift output style, reasoning patterns, and edge case handling in ways that don’t always surface on your benchmark set.


• Prompt changes can have nonlinear effects downstream, especially in multi-step agentic flows.


• Retrieval pipeline drift changes what context your agent actually sees at inference time. A document index fresh last month may have stale or subtly different content today.


• Real-world traffic distribution is never exactly what you sampled for your test set. Production surfaces long-tail inputs that feel obvious in hindsight but were invisible during development.

The implication is straightforward: evaluation has to be continuous, not episodic. You need quality signals at development time, at every CI/CD commit, and continuously against live production traffic – all using the same evaluator definitions so results are comparable across environments. That’s the core design principle behind Foundry Observability.

Continuous Evaluation Across the Full AI Lifecycle

Built-In Evaluators

Foundry’s built-in evaluators cover the most critical quality and safety dimensions for production agent systems:

• Coherence and Relevance measure whether responses are internally consistent and on-topic relative to the input. These are table-stakes signals for any conversational or task-completion agent.


• Groundedness is particularly important for RAG-based architectures. It measures whether the model’s output is actually supported by the retrieved context – as opposed to plausible-sounding content the model generated from its parametric memory. Groundedness failures are a leading indicator of hallucination risk in production, and they’re often invisible to human reviewers at scale.


• Retrieval Quality evaluates the retrieval step independently from generation. Groundedness failures can originate in two places: the model may be ignoring good context, or the retrieval pipeline may not be surfacing relevant context in the first place. Splitting these signals makes it much easier to pinpoint root cause.


• Safety and Policy Alignment evaluates whether outputs meet your deployment’s policy requirements – content safety, topic restrictions, response format compliance, and similar constraints.
  
  
  

   

  
  

These evaluators are designed to run at every stage of the AI lifecycle:

• Local development – run evals inline as you iterate on prompts, retrieval config, or orchestration logic


• CI/CD pipelines – gate every commit against your quality baselines; catch regressions before they reach production


• Production traffic monitoring – continuously evaluate sampled live traffic and surface trends over time

Because the evaluators are identical across all three contexts, a score in CI means the same thing as a score in production monitoring. See the Practical Guide to Evaluations and the Built-in Evaluators Reference for a deeper walkthrough.

Custom Evaluators – Encoding Your Own Definition of Quality

Built-in evaluators cover common signals well, but production agents often need to satisfy criteria specific to a domain, regulatory environment, or internal standard. Foundry supports two types of custom evaluators (currently in public preview):

• LLM-as-a-Judge evaluators let you configure a prompt and grading rubric, then use a language model to apply that rubric to your agent’s outputs. This is the right approach for quality dimensions that require reasoning or contextual judgment – whether a response appropriately acknowledges uncertainty, whether a customer-facing message matches your brand tone, or whether a clinical summary meets documentation standards. You write a judge prompt with a scoring scale (e.g., 1–5 with criteria for each level) that evaluates a given {input} / {response} pair. Foundry runs this at scale and aggregates scores into your dashboards alongside built-in results.


• Code-based evaluators are Python functions that implement any evaluation logic you can express programmatically – regex matching, schema validation, business rule checks, compliance assertions, or calls to external systems. If your organization has documented policies about what a valid agent response looks like, you can encode those policies directly into your evaluation pipeline.

Custom and built-in evaluators compose naturally – running against the same traffic, producing results in the same schema, feeding into the same dashboards and alert rules.

Monitoring and Alerting – AI Quality as an Operational Signal

All observability data produced by Foundry – evaluation results, traces, latency, token usage, and quality metrics – is published directly to Azure Monitor. This is where the integration pays off for teams already on Azure.

What this enables that siloed AI monitoring tools can’t:

• Cross-stack correlation. When your groundedness score drops, is it a model update, a retrieval pipeline issue, or an infrastructure problem affecting latency? With AI quality signals and infrastructure telemetry in the same Azure Monitor Application Insights workspace, you can answer that in minutes rather than hours of manual correlation across disconnected systems.


• Unified alerting. Configure Azure Monitor alert rules on any evaluation metric – trigger a PagerDuty incident when groundedness drops below threshold, send a Teams notification when safety violations spike, or create automated runbook responses when retrieval quality degrades. These are the same alert mechanisms your SRE team already uses.


• Enterprise governance by default. Azure Monitor’s RBAC, retention policies, diagnostic settings, and audit logging apply automatically to all AI observability data. You inherit the governance framework your organization has already built and approved.


• Grafana and existing dashboards. If your team uses Azure Managed Grafana, evaluation metrics can flow into existing dashboards alongside your other operational metrics – a single pane of glass for application health, infrastructure performance, and AI agent quality.

The Agent Monitoring Dashboard in the Foundry portal provides an AI-native view out of the box – evaluation metric trends, safety threshold status, quality score distributions, and latency breakdowns. Everything in that dashboard is backed by Azure Monitor data, so SRE teams can always drill deeper.

End-to-End Tracing: From Quality Signal to Root Cause

A groundedness score tells you something is wrong. A trace tells you exactly where the failure occurred and what the agent actually did.

Foundry provides OpenTelemetry-based distributed tracing that follows each request through your entire agent system: model calls, tool invocations, retrieval steps, orchestration logic, and cross-agent handoffs. Traces capture the full execution path – inputs, outputs, latency at each step, tool call parameters and responses, and token usage.

The key design decision: evaluation results are linked directly to traces. When you see a low groundedness score in your monitoring dashboard, you navigate directly to the specific trace that produced it – no manual timestamp correlation, no separate trace ID lookup. The connection is made automatically.

Foundry auto-collects traces across the frameworks your agents are likely already built on:

• Microsoft Agent Framework


• Semantic Kernel


• LangChain and LangGraph


• OpenAI Agents SDK

For custom or less common orchestration frameworks, the Azure Monitor OpenTelemetry Distro provides an instrumentation path. Microsoft is also contributing upstream to the OpenTelemetry project – working with Cisco Outshift, we’ve contributed semantic conventions for multi-agent trace correlation, standardizing how agent identity, task context, and cross-agent handoffs are represented in OTel spans.

Note: Tracing is currently in public preview, with GA shipping by end of March.

Prompt Optimizer (Public Preview)

One persistent friction point in agent development is the iteration loop between writing prompts and measuring their effect. You make a change, run your evals, look at the delta, try to infer what about the change mattered, and repeat.

Prompt Optimizer tightens this loop. It analyzes your existing prompt and applies structured prompt engineering techniques – clarifying ambiguous instructions, improving formatting for model comprehension, restructuring few-shot examples, making implicit constraints explicit – with paragraph-level explanations for every change it makes.

The transparency is deliberate. Rather than producing a black-box “optimized” prompt, it shows you exactly what it changed and why. You can add constraints, trigger another optimization pass, and iterate until satisfied. When you’re done, apply it with one click.

The value compounds alongside continuous evaluation: run your eval suite against the current prompt, optimize, run evals again, see the measured improvement. That feedback loop – optimize, measure, optimize – is the closest thing to a systematic approach to prompt engineering that currently exists.

What Makes our Approach to Observability Different

There are other evaluation and observability tools in the AI ecosystem. The differentiation in Foundry’s approach comes down to specific architectural choices:

• Unified lifecycle coverage, not just pre-deployment testing. Most existing evaluation tools are designed for offline, pre-deployment use. Foundry’s evaluators run in the same form at development time, in CI/CD, and against live production traffic. Your quality metrics are actually comparable across the lifecycle – you can tell whether production quality matches what you saw in testing, rather than operating two separate measurement systems that can’t be compared.


• No separate observability silo. Publishing all observability data to Azure Monitor means you don’t operate a separate system for AI quality alongside your existing infrastructure monitoring. AI incidents route through your existing on-call rotations. AI quality data is subject to the same retention and compliance controls as the rest of your telemetry.


• Framework-agnostic tracing. Auto-instrumentation across Semantic Kernel, LangChain, LangGraph, and the OpenAI Agents SDK means you’re not locked into a specific orchestration framework. The OpenTelemetry foundation means trace data is portable to any compatible backend, protecting your investment as the tooling landscape evolves.


• Composable evaluators. Built-in and custom evaluators run in the same pipeline, against the same traffic, producing results in the same schema, feeding into the same dashboards and alert rules. You don’t choose between generic coverage and domain-specific precision – you get both.


• Evaluation linked to traces. Most systems treat evaluation and tracing as separate concerns. Foundry treats them as two views of the same event – closing the loop between detecting a quality problem and diagnosing it.

Getting Started

If you’re building agents on Microsoft Foundry, or using Semantic Kernel, LangChain, LangGraph, or the OpenAI Agents SDK and want to add production observability, the entry point is Foundry Control Plane.