AI Wisdom

NVIDIA professional level

NVIDIA Certified Professional: Agentic AI

Code
NCP-AAI
Length
120 minutes
Price
$200
Valid for
2 years, renewed by retaking the exam
Study time, our estimate
80 to 120 hours

A broad agents exam, and more vendor-neutral than you might expect. Only 7% is about the NVIDIA platform. The rest is agent design, memory, evaluation and deployment. Useful if you want an agents credential that is not tied to a cloud.

Take it if

You build production agents and want a credential that is not tied to one cloud.

Skip it if

You have not shipped an agent yet. NVIDIA expects 1 to 2 years of experience, with production projects.

Before you start

Check yourself against this list first. If something is missing, close that gap before you pay for an exam.

  • One to two years in AI/MLNVIDIA lists this.
  • Production agent projectsShip one first.

Helpful first: NCA-GENL

All exam details
Format
60 to 70 questions
Length
120 minutes
Price
$200
Valid for
2 years, renewed by retaking the exam
Delivery
Proctored, English
Experience
1 to 2 years in AI/ML, with production agent projects
Status
NVIDIA's page shows both "Coming soon" and a register button, so confirm before you plan around it
If you fail
Wait 14 days

What the exam covers

  • Agent Architecture and Design15%
  • Agent Development15%
  • Evaluation and Tuning13%
  • Deployment and Scaling13%
  • Cognition, Planning, and Memory10%
  • Knowledge Integration and Data Handling10%
  • NVIDIA Platform Implementation7%
  • Run, Monitor, and Maintain5%
  • Safety, Ethics, and Compliance5%
  • Human-AI Interaction and Oversight5%

What to focus on

  • Agent architecture and development make up nearly a third of the exam.
  • Planning, memory, and how agents reason across steps.
  • Evaluation and tuning, at 13%.
  • Deployment and scaling, at 13%.
  • Knowledge integration: RAG inside agents.
  • Human oversight and safety.
Everything on one mapEvery topic and key idea in this certification.
mindmap
  root((NCP-AAI))
    Architecture and development
      When to use agents
      Tool schemas
      Frameworks
    Planning, memory and knowledge
      Replanning
      Memory
      Knowledge
    Evaluation
      Task success
      Tool accuracy
      Cost
    Deploy and operate
      Stateless services
      NIM
      Monitoring
    Safety and oversight
      Human approval
      Guardrails
      Accountability

Lesson 1 of 5, 15% + 15% of the exam

Architecture and development

Thirty percent of the exam.

When to use agents
When the path depends on what is found.
Tool schemas
Clear schemas plus validation reduce bad calls.
Frameworks
NVIDIA documents NeMo Agent Toolkit, which it deprecated in October 2026.
Anatomy of an agentThe parts the exam expects you to name.
classDiagram
  class Agent {
    planner
    reasoning
  }
  class Tools {
    schema
    validation
  }
  class Memory {
    short term
    long term
  }
  class Knowledge {
    RAG
  }
  Agent --> Tools
  Agent --> Memory
  Agent --> Knowledge

Common trap

Relying on older NVIDIA agent tooling. Check its status before you build on it.

Try it yourself

Build one agent with two tools and validation.

Official docs: NeMo Agent Toolkit

Lesson 2 of 5, 10% + 10% of the exam

Planning, memory and knowledge

This is what makes an agent more than a chatbot.

Replanning
Adjust the plan on new information.
Memory
Context lasts a session. Durable facts need external storage.
Knowledge
Bring RAG in as a tool.
A planning loopThe agent reacts to what it observes.
stateDiagram-v2
  direction LR
  [*] --> Plan
  Plan --> Act
  Act --> Observe
  Observe --> Plan: Unexpected result
  Observe --> Act: On track
  Observe --> Done: Goal met
  Observe --> Human: Stuck or risky
  Done --> [*]

Common trap

Retrying the same failed step forever.

Try it yourself

Add long-term memory to an agent.

Official docs: NVIDIA training

Lesson 3 of 5, 13% of the exam

Evaluation

Agents fail in ways unit tests miss.

Task success
Measure on a fixed scenario set.
Tool accuracy
Did it call the right tool correctly?
Cost
Track tokens and steps per task.
Scoring an agentOne number hides the trade-offs.
mindmap
  root((Agent quality))
    Task success
    Tool accuracy
    Steps taken
    Cost per task
    Safety

Common trap

Using one number to judge an agent.

Try it yourself

Write ten scenarios and score two agent versions.

Official docs: NVIDIA training

Lesson 4 of 5, 13% + 7% + 5% of the exam

Deploy and operate

Scaling agents differs from scaling a model.

Stateless services
Keep state external so you can add replicas.
NIM
Serves the models behind the agent.
Monitoring
Trace steps and tool calls.
Scaling agent servicesState moves out, so replicas are interchangeable.
flowchart LR
  A["Load balancer"] --> B["Agent replica 1"]
  A --> C["Agent replica 2"]
  B --> D[("Shared state")]
  C --> D
  B --> E["Model service"]
  C --> E

Common trap

Storing session state in process memory.

Try it yourself

Run two replicas of an agent with shared state.

Official docs: NVIDIA NIM

Lesson 5 of 5, 5% + 5% of the exam

Safety and oversight

Ten percent of the exam.

Human approval
Required for irreversible actions.
Guardrails
Block unsafe inputs and outputs.
Accountability
Keep a record of decisions.
Oversight by riskA person steps in where actions cannot be undone.
stateDiagram-v2
  direction LR
  [*] --> Propose
  Propose --> Execute: Low risk
  Propose --> Human: High risk
  Human --> Execute: Approved
  Human --> [*]: Rejected
  Execute --> [*]

Common trap

Relying on a warning in the system prompt.

Try it yourself

Add an approval step before one risky tool.

Official docs: NeMo Guardrails

Study plan

  1. Weeks 1 and 2: RAG agents

    Take Building RAG Agents With LLMs, then Evaluating RAG and Semantic Search Systems.

  2. Week 3: agents

    Take Building Agentic AI Applications With LLMs and build a planner with memory.

  3. Week 4: deploy

    Take Introduction to Deploying RAG Pipelines for Production at Scale. Add monitoring and a human approval step.

  4. Exam week

    Take the mock below and restudy any domain you missed.

Exam day and after

  • NVIDIA certifications last two years, and you renew by retaking the exam.
  • Exams are proctored and in English. NVIDIA's FAQ covers both remote and in-person testing, and remote exams allow no breaks.
  • Check the exam page before you plan around a date. NVIDIA has been adding hands-on labs to professional exams.
  • The published weights add up to 98%, so a small part of the exam is not accounted for.
  • NVIDIA deprecated NeMo Agent Toolkit in October 2026, with end of life by the end of 2026. Treat study material built on it with care.
Test yourself8 practice questions, scored by exam topic.
Start the mock exam

Keep going

Exam facts checked on October 31, 2026 against NVIDIA certification page. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.