Independent: not affiliated with NVIDIA. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its NVIDIA source.
5 questions / Explanations link NVIDIA sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Group 1 / Agent architecture and design
A media company's nightly job uses the NeMo Agent Toolkit's Sequential Executor to run three type-annotated functions on each new podcast episode: transcribe the audio, mark chapters, write show notes. A developer changed the chapter function to return a list of chapter objects, but the show-notes function still takes a string. The next night's run logged a type-compatibility warning, carried on, and published garbled show notes. The team wants a mismatch like this to stop the job before any function runs. What should they change in the executor's configuration?
Show answerHide answer
Answer: D
Why
Before running the tool list, the Sequential Executor compares each function's output type with the next function's input type from their annotations. raise_type_incompatibility defaults to false, which turns a failed comparison into a warning; set to true, any incompatibility raises an exception before the first function runs.
Why not the others
AThis setting covers a tool that fails while the chain runs: the executor returns an error message as the workflow output instead of raising. The type check would still only warn, and in this run no tool failed.
BStreaming changes how the chapter function delivers its output; a failed type check still only produces a warning while raise_type_incompatibility is false.
CThe description is the tool description shown to another agent when the executor is configured as a function; no check reads it.
Group 2 / Development, cognition, planning and memory
An internal IT helpdesk assistant is built with the NeMo Agent Toolkit as a ReAct agent. It keeps long-term memory in Redis and has the add_memory and get_memory tools, so that it can remember each employee's device, office and open tickets between conversations. In review, the team finds memories saved for only some conversations: on many turns the model ends its reply without calling either tool, and the next conversation starts from nothing. The team wants every turn saved and retrieved later, not only the turns where the model chooses to use a tool. What should they change?
Show answerHide answer
Answer: A
Why
The toolkit names the weakness of tool-based memory: LLMs may forget to call memory tools, so capture is inconsistent. The auto_memory_agent workflow wraps an existing agent (inner_agent_name) and a configured memory (memory_name), stores messages and responses automatically and injects relevant memory before each call, with Redis and any agent type.
Why not the others
Btool_calling_agent changes how the model emits tool calls, not who decides to make them; the model can still end a turn without calling add_memory or get_memory.
CA prompt rule is advice to the model: on any turn it can ignore the rule, and nothing in the toolkit then writes or reads memory.
DTemperature controls how varied the sampled output is; it does not put a memory call into every turn, so skipped calls remain possible.
A support-chat workflow built with the NeMo Agent Toolkit calls an LLM served by NIM. The team has an evaluation dataset with expected answers and two evaluators, one scoring answer accuracy and one measuring latency. For three weeks engineers have changed temperature, top_p and max_tokens by hand between evaluation runs, and nobody can show which combination gives the best balance of accuracy and latency. What should the team do?
Show answerHide answer
Answer: B
Why
nat optimize automates the search: the optimizer section ties each eval_metrics entry to an evaluator with a direction and a weight, Optuna suggests values from each setting's search space (the NIM LLM type has defaults for temperature, top_p and max_tokens), and the metrics combine into one score.
Why not the others
A--reps repeats the evaluation of one configuration to average out run-to-run variation; it never changes temperature, top_p or max_tokens, so it cannot show which combination balances accuracy and latency.
COnly the evaluators listed in eval_metrics are scored during the search; with accuracy alone, latency never enters the choice, so the search can settle on a slow combination and shows no accuracy-latency trade-off.
DThe profiler records tokens, timing and LLM calls for the runs it observes; it does not change temperature, top_p or max_tokens, and choosing by latency alone ignores accuracy.
A team serves its support agent from a staging cluster with nat serve; staging has its own retrieval index and tool settings. Before each release, an engineer runs nat eval on a laptop with the release branch's workflow configuration and the team's dataset of questions and expected answers. Release 14 scored the same as release 13 in those runs, yet reviewers found it answering worse once it was deployed to staging. The team wants each release scored as deployed in staging, with the same dataset and evaluators. What should they change?
Show answerHide answer
Answer: C
Why
With --endpoint, nat eval runs the dataset's questions through the workflow on the remote server and evaluates the answers locally, so the answers come from staging's own configuration, index and tools while the same dataset and evaluators score them. Enabling append_job_id_to_output_dir keeps each run's files in its own folder.
Why not the others
AThe --endpoint flag works, but the server it reaches would run the laptop's copy of the release configuration, so the scores still would not describe staging.
BA base_url change only redirects the model calls; the workflow's prompts, tools and retrieval settings would still come from the laptop's configuration, not from staging.
DRepetitions steady the laptop's scores; they do not make the laptop's workflow match the one running in staging.
A support agent built with the NeMo Agent Toolkit updates customer records. When a caller's name matches three records, the agent must ask the support rep which record to update and then continue in the same run. The workflow is served with nat serve and used through the toolkit's web UI. How should the record-update function get the rep's choice?
Show answerHide answer
Answer: D
Why
A function gets the UserInteractionManager from the Context and calls prompt_user_input; the run pauses until the user responds. A radio prompt takes a list of options and returns a HumanResponseRadio whose selected_option holds the chosen record. These prompts work under nat serve; the web UI needs WebSocket mode on.
Why not the others
AA binary prompt carries exactly two options, and the toolkit rejects any other number, so it cannot list three records.
BA notification prompt is for informing the rep; its response model is an acknowledgement (the default text is 'Notification acknowledgement.'), not a selection from a list.
CA radio prompt returns a HumanResponseRadio, which has a selected_option field and no text field, so reading response.content.text, as the guide's text-prompt example does, fails.
The study bundle has two practice exams with 130 original questions, an answer key and an explanation for every answer, plus the 60-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
130 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.