NVIDIANVIDIA-Certified Professional: Agentic AI2026-100 Prerequisites
Choose who controls the next step
Completed by the knowledge check
Learning objective
Describe the agent loop and contrast agentic with predetermined workflows.
Core idea
The defining control-flow question is who selects the next operation. In an agent loop, a model uses the task and observations to select an action or return an answer. In a predetermined workflow, application logic fixes the sequence or branches, even when individual steps call a language model.
Autonomy is therefore a property of a particular decision. A system can use a fixed outer workflow and delegate one investigative stage to an agent without handing every business operation to model-selected control.
How it works
An agent receives a task and available context, selects an action, and requests a tool. The runtime executes that request and returns an observation. The model uses the observation to decide whether another action is useful or whether it can answer. A failed lookup can change the next action, rather than merely advancing to the next fixed stage.
NeMo Agent Toolkit’s ReAct agent implements this repeated reasoning, action, and observation pattern. Its _type is react_agent, configured as the workflow or as a function; llm_name identifies the configured model, and tool_names identifies the available functions or function groups. This agent requires the nvidia-nat[langchain] plugin. See 1.2.2 for the detailed loop and termination controls, and 1.2.3 for alternatives to ReAct.
The runtime also needs an execution limit. The ReAct configuration exposes max_tool_calls to bound tool use. That bound contains repeated actions; it does not prove that the final answer is correct. The application must treat exhausted work as an outcome requiring handling, rather than as successful task completion.
A predetermined counterpart is the toolkit’s sequential_executor. Its tool_list orders functions, and each function’s output becomes the next function’s input. The executor controls the chain without an LLM choosing which function comes next. A summarization function inside that chain still performs generation, but the chain’s orchestration remains predetermined. See 1.6.3 for state graphs and predetermined versus adaptive control.
In an agent loop the model selects the next operation; in a predetermined workflow the executor controls the chain.
Key distinctions and trade-offs
Generation versus agency: Generating variable text does not make a pipeline agentic. The relevant distinction is whether model output selects subsequent actions.
Adaptation versus bounded execution: An agent can pursue an unexpected observation, but additional decisions and calls spend tokens, time, and tool capacity. A fixed sequence is easier to enumerate when the required steps are already known.
Workflow versus agent container: In NeMo Agent Toolkit, workflow is also the configuration entry for a top-level execution component. The key’s name does not tell you whether its _type uses adaptive action selection.
Architecture / decision scenario
An employee IT service-desk assistant processes 5,000 tickets daily. Every ticket must pass intake, an entitlement check, and an internal summary. Only 8% require investigation across logs and change records. The service-desk team permits at most four investigative tool calls per ticket.
The tempting proposal is to give a ReAct agent all stages and let it discover the sequence. The workflow owner keeps the three mandatory stages in a sequential_executor chain. Application logic outside that chain branches on eligibility and invokes an investigative react_agent configured under functions, with max_tool_calls: 4; the sequential executor does not supply the conditional branch. See 1.6.3 for modelling that control. The agent can choose which log or record to inspect next; it cannot replace the mandatory entitlement check. This concentrates adaptive decision cost on the 8% that need it.
Common exam traps
- Calling every LLM pipeline an agent.
- Assuming a tool response ends the task rather than informing another decision.
- Treating a tool-call cap as an accuracy guarantee.
- Reading the
workflowconfiguration key as proof of predetermined control.
What to remember
Locate the decision that selects the next operation. Use fixed control for known sequences and bounded agent control where observations must change the path. A hybrid system can contain both.
Knowledge check
In NeMo Agent Toolkit, a team configures the top-level workflow as a sequential_executor whose tool_list runs a claim summarizer, which calls an LLM, and then a report writer. Which description of this workflow’s control flow is accurate?
workflow keyworkflow key names the top-level component and says nothing about its control flow.A NeMo Agent Toolkit react_agent that investigates billing disputes is configured with max_tool_calls: 4. On some disputes it reaches that bound before it has an answer it can support. How should the application treat those runs?
max_tool_calls contains repeated actions; it does not make a run’s result correct or complete, so a run that stops without a supported answer remains unfinished even though it stayed within budget. A larger budget can permit further investigation, but returning an answer still does not establish its correctness.A claims service processes 10,000 claims a day. Every claim must pass document intake, a coverage check and a ledger entry, in that order. About 5% are flagged for fraud review, which follows leads across payment and policy records, where each next lookup depends on what the last one returned. Which design limits adaptive control to the work that needs it?
References
- NVIDIA, Configure the ReAct Agent
- NVIDIA, Configure Sequential Executor