Skip to lesson

NVIDIANVIDIA-Certified Professional: Agentic AI2026-100 Prerequisites

Separate availability, quality, telemetry, and authority

Completed by the knowledge check

Learning objective

Explain serving, evaluation, telemetry, and application security basics.

Core idea

A reachable model endpoint proves little about whether an agent completes the right task safely. Production readiness requires separate answers to four questions: can the system serve requests, does it produce acceptable outcomes, can operators explain its behavior, and which actions is it authorized to take?

These questions need different measurements and controls. A successful HTTP response is not a quality score, and a quality score is not permission to read another employee’s records.

How it works

Serving exposes inference to callers and operates the runtime that performs it. NIM for LLM and VLM separates /v1/health/live, which indicates container liveness, from /v1/health/ready, which indicates the model is loaded and available for inference. These signals establish runtime availability, not answer correctness. See 7.2.1 for deployment health and 8.1.1 for service-level objectives and reliability metrics.

Evaluation compares outputs or trajectories with task-specific expectations. In NeMo Agent Toolkit, nat eval executes a configured workflow on a dataset and runs configured evaluators. The eval.general.dataset section identifies inputs, and eval.evaluators selects the scoring components. The full config-driven runtime requires nvidia-nat[eval]. The essential prerequisite is a test set with defined success criteria, not a dashboard of successful HTTP requests. See 3.1.1 for dataset design and 3.1.4 for repeatable runs.

Telemetry records runtime behavior. OpenTelemetry distinguishes traces, the path of a request through the application; metrics, runtime measurements; and logs, event records. Use each for its appropriate question: a latency metric reveals a trend, while a trace can locate which part of a request consumed time. See 8.2.1 for correlating spans, structured logs, and request IDs. Telemetry helps diagnose an outcome; it does not independently label that outcome correct.

Application security constrains access and action. OWASP’s Excessive Agency guidance distinguishes excessive functionality, permissions, and autonomy. Its concrete controls include narrowing tool functions, using the requesting user’s authorization context, and enforcing minimal permissions in downstream systems. A database identity with SELECT on the required table is different from one with UPDATE or DELETE. See 9.1.1 for least privilege and credential isolation. NeMo Guardrails’ input and output rails add content checks; see 7.1.1 for rail types. Content checks and downstream authorization address different failure paths.

Key distinctions and trade-offs

Evaluation spends execution and scoring resources to estimate quality on selected cases. Telemetry describes actual operation, including cases your test set may omit. Neither replaces the other.

Security also needs enforcement at the system that holds the resource. A prompt asking a model to respect permissions cannot remove privileges from its database identity. Broader tool access may make more tasks possible while increasing the consequences of a mistaken action.

Architecture / decision scenario

A healthcare employee assistant returns HTTP 200 on 99.9% of requests, but a labeled set shows only 86% correct policy answers against a 95% release requirement. Traces also show a six-second retrieval step against a four-second end-to-end latency target. Its proposed database identity can read every department’s records.

The product owner rejects the tempting conclusion that high availability means release readiness. The evaluation owner investigates the failed answer cases; the operations owner uses the trace to address retrieval delay; the security owner restricts database access to the requesting employee’s permitted records. Raising replicas alone resolves neither the measured quality gap nor excessive authority. All figures are scenario requirements or observations, not vendor guarantees.

Common exam traps

  • Substituting endpoint health for task success.
  • Treating logs as a labeled evaluation dataset.
  • Using a content filter as the authorization system.
  • Granting broad credentials because a prompt forbids misuse.

What to remember

Serving supplies availability, evaluation measures task quality, telemetry explains operation, and security limits authority. Assess each explicitly before treating an agent as ready for production.

Knowledge check

References