I85PREP
⌘ K
Oracle Agentic AI Foundations study materials
Oracle Agentic AI Foundations study resources

Oracle Agentic AI Foundations practice questions

One question at a time. Build your Oracle Agentic AI Foundations understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Agentic AI Concepts and Architecture

In an agentic AI system, which component is responsible for deciding the next action based on the current state and the agent's objective?

  1. The observability exporter
  2. The vector store
  3. The planning module Correct answer.
  4. The embedding model

Explanation

The planning module (often driven by an LLM reasoning loop) interprets the goal and current state, then selects the next step or tool call. Vector stores and embedding models support retrieval, while observability exporters only record telemetry.

Question 2Agentic AI Concepts and Architecture

Which characteristic most distinguishes an agentic workflow from a single-shot LLM prompt?

  1. It removes the need for any external tools
  2. It guarantees deterministic outputs
  3. It iterates through multiple reasoning and action steps toward a goal Correct answer.
  4. It always uses a larger context window

Explanation

Agentic behavior means the system loops through observe-reason-act cycles, using intermediate results to decide subsequent steps. Context size, tool use, and determinism are separate concerns and not defining properties.

Question 3Agentic AI Concepts and Architecture

A support agent must look up a customer's order status in a live database before replying. Which mechanism enables this in an agent architecture?

  1. Fine-tuning the base model on order data
  2. Compressing the conversation history
  3. Increasing the temperature setting
  4. Tool calling invoked by the agent's reasoning loop Correct answer.

Explanation

Tool calling lets the agent invoke an external function or API, receive structured results, and incorporate them into the next reasoning step. Fine-tuning bakes in patterns but cannot fetch live order state.

Question 4Agentic AI Concepts and Architecture

What is the primary purpose of a ReAct-style loop in an agent?

  1. To reduce token cost by caching prompts
  2. To interleave reasoning traces with actions and observations Correct answer.
  3. To enforce strict JSON output formatting
  4. To train the underlying model on new data

Explanation

ReAct alternates between generating reasoning (thought), taking an action, and observing the result, which grounds later reasoning in real feedback. It is not a formatting, caching, or training technique.

Question 5Agentic AI Concepts and Architecture

In agent terminology, what is a 'tool'?

  1. An external function or API the agent can call to perform an action Correct answer.
  2. A dashboard for monitoring agent runs
  3. A fine-tuned model checkpoint
  4. A configuration file that stores API keys

Explanation

Tools are callable capabilities, such as search, calculators, or database queries, exposed to the agent with a described interface. The agent selects a tool, supplies arguments, and receives a result.

Question 6Agentic AI Concepts and Architecture

Which memory type lets an agent recall facts from earlier in the same conversation?

  1. Episodic memory across sessions
  2. Procedural memory of tool schemas
  3. Short-term conversational memory Correct answer.
  4. Semantic memory in a vector store

Explanation

Short-term memory holds the current session's message history so the agent can reference earlier turns. Semantic, episodic, and procedural memory are longer-lived stores with different purposes.

Question 7Agentic AI Concepts and Architecture

What does the 'observation' step in an agent loop typically contain?

  1. The system prompt instructions
  2. The agent's internal goal statement
  3. The output returned by the tool or environment after an action Correct answer.
  4. The list of available tool names

Explanation

An observation is the feedback from the environment or tool result that follows an action, and it informs the next reasoning step. Goals, system prompts, and tool catalogs are inputs, not observations.

Question 8Agentic AI Concepts and Architecture

Why is grounding an agent's responses in retrieved documents important?

  1. It reduces fabricated claims by tying answers to verifiable sources Correct answer.
  2. It eliminates the need for any system prompt
  3. It guarantees the agent never makes an incorrect tool call
  4. It makes the model's weights smaller

Explanation

Retrieval grounding supplies evidence the model can cite, lowering hallucination risk. It does not shrink model weights, replace prompts, or guarantee correct tool selection.

Question 9Agentic AI Concepts and Architecture

An agent repeatedly calls the same search tool with an identical query and never progresses. Which architectural control most directly addresses this?

  1. A larger embedding dimension
  2. A longer system prompt
  3. A higher top-k retrieval value
  4. A loop-detection guard that limits repeated identical actions Correct answer.

Explanation

Loop detection tracks recent actions and can break cycles by forcing a different strategy or terminating. Embedding size, retrieval depth, and prompt length do not prevent repeated identical calls.

Question 10Agentic AI Concepts and Architecture

Which pattern best describes a supervisor agent coordinating several specialized worker agents?

  1. Single-shot prompting
  2. Model quantization
  3. Multi-agent orchestration Correct answer.
  4. Retrieval-augmented generation

Explanation

Multi-agent orchestration assigns a coordinator to decompose the task and route subtasks to specialized agents, then aggregate results. The other options are prompting, retrieval, and model-compression techniques.

Question 11Agentic AI Concepts and Architecture

What is the main benefit of giving an agent a structured tool schema with typed parameters?

  1. It lets the runtime validate arguments before executing the call Correct answer.
  2. It removes the need for error handling
  3. It guarantees the tool returns correct results
  4. It trains the model on new tool data

Explanation

Typed schemas allow validation and clear prompting of parameter formats, reducing malformed calls. They do not train the model, replace error handling, or guarantee correct tool outputs.

Question 12Agentic AI Concepts and Architecture

A team wants the agent to recover when a tool returns a transient error. Which design is most appropriate?

  1. Silently return an empty string to the user
  2. Disable the tool permanently
  3. Retry with backoff and surface the failure to the reasoning loop Correct answer.
  4. Increase the model temperature

Explanation

Retrying with backoff handles transient faults, and passing the failure back as an observation lets the agent choose an alternative. Hiding errors or disabling tools degrades reliability.

Question 13Agentic AI Concepts and Architecture

Which statement best captures the role of a system prompt in an agent architecture?

  1. It stores the long-term vector index
  2. It replaces the need for observations
  3. It executes the tool calls directly
  4. It defines the agent's role, constraints, and operating procedure Correct answer.

Explanation

The system prompt establishes persona, boundaries, and procedural guidance that shape every reasoning step. It does not store indexes, execute tools, or substitute for environment feedback.

Question 14Agentic AI Concepts and Architecture

Why might an architect add a human-in-the-loop approval step before an agent executes a high-impact action?

  1. To eliminate the need for logging
  2. To reduce the model's parameter count
  3. To provide oversight and prevent unintended irreversible changes Correct answer.
  4. To speed up every tool call

Explanation

Human approval gates risky actions such as payments or deletions, adding oversight before irreversible effects occur. It does not affect model size, speed up calls, or replace logging.

Question 15Agentic AI Concepts and Architecture

What is the function of an orchestrator in an agentic system?

  1. It renders the user interface
  2. It hosts the model weights on a GPU
  3. It routes tasks, manages state, and sequences agent steps Correct answer.
  4. It converts documents into embeddings

Explanation

The orchestrator coordinates the control flow, tracks state, and dispatches work to models, tools, or sub-agents. Embedding, hosting, and UI rendering are separate infrastructure roles.

Question 16Agentic AI Concepts and Architecture

An agent must answer questions over a large private document set. Which architecture component is most directly needed?

  1. A second planning agent with no tools
  2. A larger output token limit
  3. A retrieval layer with a vector index over the documents Correct answer.
  4. A higher temperature setting

Explanation

A retrieval layer chunks and indexes documents so relevant passages can be supplied to the model at query time. Planning agents without tools, token limits, and temperature do not provide document access.

Question 17Agentic AI Concepts and Architecture

Which is the clearest example of an agent's episodic memory?

  1. A record of a past session where a specific refund was processed Correct answer.
  2. The current system prompt
  3. A stored schema describing a payment tool
  4. A cached embedding of the product catalog

Explanation

Episodic memory stores records of specific past events or sessions that can be recalled later. Tool schemas are procedural knowledge, embeddings are semantic artifacts, and the system prompt is configuration.

Question 18Agentic AI Concepts and Architecture

What does 'task decomposition' mean in agent planning?

  1. Breaking a complex goal into smaller executable subtasks Correct answer.
  2. Compressing the prompt to save tokens
  3. Reducing the model's parameter count
  4. Splitting a dataset for training and testing

Explanation

Task decomposition splits a high-level objective into manageable steps that can be planned, executed, and verified individually. The other options concern model size, data splits, and prompt compression.

Question 19Agentic AI Concepts and Architecture

An agent's tool call returns a JSON error indicating an invalid argument. What should the agent do next to be robust?

  1. Ignore the error and fabricate a plausible result
  2. Switch to a different language model
  3. Read the error, correct the arguments, and retry the call Correct answer.
  4. Terminate the whole application immediately

Explanation

Treating the error as an observation lets the agent diagnose the argument problem and retry with corrected input. Fabricating results or crashing harms reliability, and switching models is unrelated.

Question 20Agentic AI Concepts and Architecture

Which design most reduces the risk of an agent taking an unsafe action in production?

  1. Removing all logging to save storage
  2. Least-privilege tool permissions scoped to the task Correct answer.
  3. Giving the agent admin credentials for convenience
  4. Allowing unlimited retries without limits

Explanation

Least-privilege permissions limit the blast radius if the agent misbehaves, restricting tools to what the task requires. Broad credentials, absent logging, and unbounded retries all increase risk.

Question 21Agentic AI Concepts and Architecture

An agent's plan succeeds on simple tasks but fails on multi-step tasks requiring intermediate verification. Which architectural change best targets this?

  1. Remove tool descriptions from the prompt
  2. Lower the retrieval top-k
  3. Increase the model's context window only
  4. Add explicit reflection and self-check steps between subtasks Correct answer.

Explanation

Reflection steps let the agent evaluate intermediate results and correct course before proceeding, which directly addresses multi-step reliability. Context size alone, retrieval depth, and removing tool descriptions do not add verification.

Question 22Agentic AI Concepts and Architecture

A multi-agent system shows agents duplicating work and conflicting on shared state. Which architectural fix is most appropriate?

  1. Remove all inter-agent messaging
  2. Give every agent write access to all shared resources
  3. Introduce a coordinator that owns task assignment and state transitions Correct answer.
  4. Increase each agent's temperature

Explanation

A coordinator centralizes assignment and state ownership, preventing duplicate effort and conflicting writes. Broad write access worsens conflicts, and removing messaging or raising temperature does not resolve coordination.

Question 23Agentic AI Concepts and Architecture

Which scenario best demonstrates the difference between semantic and procedural memory in an agent?

  1. Semantic memory stores API keys, while procedural memory stores chat history
  2. Semantic memory stores session logs, while procedural memory stores embeddings
  3. Semantic memory stores tool outputs, while procedural memory stores the system prompt
  4. Semantic memory stores product facts, while procedural memory stores how to run a refund workflow Correct answer.

Explanation

Semantic memory holds declarative knowledge such as facts, while procedural memory holds know-how such as workflows and tool usage patterns. The other pairings mislabel logs, prompts, keys, or history.

Question 24Agentic AI Concepts and Architecture

An agent must choose among several tools with overlapping descriptions. Which practice most improves reliable tool selection?

  1. Using identical descriptions for similar tools
  2. Adding more tools with vague descriptions
  3. Hiding parameter types from the model
  4. Writing distinct, narrowly scoped tool descriptions with clear parameter semantics Correct answer.

Explanation

Distinct, narrowly scoped descriptions reduce ambiguity so the model can match intent to the correct tool. Vague or duplicated descriptions, and hidden parameter types, increase misselection.

Question 25Agentic AI Concepts and Architecture

When evaluating an agent, why is measuring end-to-end task success rate more informative than measuring single-step accuracy alone?

  1. It eliminates the need for per-step tracing
  2. It captures compounding errors and coordination across the full reasoning loop Correct answer.
  3. It guarantees the agent has no hallucinations
  4. It measures only model parameter size

Explanation

End-to-end success reflects how errors compound and how well steps coordinate, which single-step accuracy can hide. It does not guarantee zero hallucination, replace tracing, or measure parameter size.

Question 26Agentic AI Concepts and Architecture

In an agentic loop, what does the 'reasoning' (or thought) step primarily produce before any tool is invoked?

  1. A signed audit record written to the orchestration log
  2. A cached copy of the tool's previous response for reuse
  3. A compressed vector representation of the user's original query
  4. A natural-language or structured rationale that decides the next action and selects a tool Correct answer.

Explanation

The reasoning step is where the model interprets the current goal and observation, then decides what to do next, including whether and which tool to call. It is an internal decision artifact, not a persistence or logging mechanism, and it precedes the action rather than following it.

Question 27Agentic AI Concepts and Architecture

Which characteristic most clearly distinguishes an agentic workflow from a single-shot LLM prompt chain?

  1. It always routes requests through a retrieval index first
  2. It iteratively selects actions and observes results until a goal condition is met Correct answer.
  3. It returns responses in a fixed JSON schema
  4. It uses a larger context window for each request

Explanation

Agentic behavior is defined by a control loop in which the system chooses actions, observes outcomes, and repeats until the goal is satisfied or a stopping condition triggers. Larger contexts, retrieval, and schema formatting can appear in non-agentic pipelines too, so they do not define agency.

Question 28Agentic AI Concepts and Architecture

An architect wants the agent to stop cleanly instead of looping forever when a task cannot be completed. Which control is most directly responsible for this?

  1. Storing the full transcript in long-term memory
  2. Adding an explicit termination condition with a maximum iteration budget Correct answer.
  3. Increasing the model's temperature for more varied attempts
  4. Expanding the tool schema with additional optional parameters

Explanation

A bounded loop requires a declared success criterion plus a hard cap on iterations or time so the agent exits even when the goal is unreachable. Without such a guard the loop can continue indefinitely regardless of model creativity or memory size.

Question 29Agentic AI Concepts and Architecture

Which component is responsible for translating a model's requested tool invocation into an actual authenticated API call and returning the result to the agent?

  1. The planner module that decomposes the goal
  2. The guardrail policy that filters outputs
  3. The memory store that holds prior turns
  4. The tool executor or integration layer Correct answer.

Explanation

The tool executor mediates between the model's abstract request and the real external system, handling authentication, argument validation, and result formatting. Planning decides what to call, memory retains context, and guardrails constrain behavior, but none of them performs the invocation.

Question 30Agentic AI Concepts and Architecture

A support agent must look up order status, then issue a refund only if the order is eligible. Which architecture element most directly enforces the eligibility gate before the refund tool can run?

  1. A deterministic policy check in the orchestration layer that conditions the refund tool on the lookup result Correct answer.
  2. A second language model that reviews the final answer's tone
  3. An embedding index of the refund policy document
  4. A longer system prompt instructing the model to be careful about refunds

Explanation

Placing the eligibility rule as deterministic orchestration logic guarantees the refund tool cannot execute unless the lookup returns an eligible status, independent of model phrasing. Prompt instructions and retrieved policy text inform the model but do not hard-enforce the precondition.

Question 31Agentic AI Concepts and Architecture

An agent's context window is filling with raw tool outputs during a long task. Which architectural technique best preserves relevant information while keeping the context bounded?

  1. Repeating the system prompt after every tool call
  2. Disabling tools once the context reaches a threshold
  3. Raising the model temperature to shorten responses
  4. Summarizing or compressing older observations and retaining key state in external memory Correct answer.

Explanation

Context management typically combines summarization of older turns with an external store for durable state, so the working context stays within limits while critical facts remain retrievable. Simply cutting off tools or repeating instructions does not reduce the volume of accumulated observations.

Question 32Agentic AI Concepts and Architecture

A team wants the agent to plan before acting on complex requests. Which design most directly supports explicit, inspectable planning?

  1. Relying on the model's latent reasoning without surfacing intermediate steps
  2. Having the agent emit a structured task list that is executed and revised step by step Correct answer.
  3. Allowing the agent to call tools in parallel without a declared sequence
  4. Removing the observation step to shorten the loop

Explanation

An explicit, structured plan makes the decomposition visible and reviewable, and it can be revised as observations arrive. Parallel unplanned calls, hidden reasoning, and omitted observations all reduce the ability to inspect and correct the plan.

Question 33Agentic AI Concepts and Architecture

Which statement best describes the difference between a workflow and an agent in terms of control flow?

  1. A workflow always uses multiple models, while an agent uses exactly one
  2. A workflow cannot use an LLM, while an agent must
  3. An agent cannot call external tools, while a workflow can
  4. A workflow's path is largely predetermined by code, while an agent's path is chosen at runtime by the model Correct answer.

Explanation

The defining contrast is where control flow is decided: workflows encode the sequence in application logic, whereas agents let the model select the next step dynamically based on observations. Model count and tool access are implementation details that either pattern may share.

Question 34Agentic AI Concepts and Architecture

A retrieval-augmented agent answers a question using a document chunk that is outdated. Which architectural change most directly addresses stale knowledge?

  1. Lowering the model temperature for the final answer
  2. Shortening the system prompt to save tokens
  3. Increasing the number of retrieved chunks per query
  4. Adding metadata filters and freshness ranking to the retrieval step Correct answer.

Explanation

Freshness problems are solved at retrieval by filtering or ranking on document metadata such as date or version so current sources are preferred. Retrieving more chunks or changing sampling settings does not correct the underlying staleness of the sources.

Question 35Agentic AI Concepts and Architecture

An agent is given a tool that can send emails. Which safeguard most directly limits the blast radius of a mistaken invocation?

  1. Logging the email content after it has been sent
  2. Increasing the model's context window so it sees more history
  3. Requiring the recipient and content to pass a validation and approval step before dispatch Correct answer.
  4. Adding a second tool that can also send emails

Explanation

Blast radius is controlled by intercepting the action before it takes effect, for example through validation and human or policy approval of recipient and content. Post-hoc logging and additional sending tools increase exposure rather than constrain it.

Question 36Agentic AI Concepts and Architecture

Which memory design best supports an agent that must remember a user's stable preferences across many separate sessions?

  1. Keeping the full chat transcript in the current context window
  2. Writing durable user-profile facts to a persistent store keyed by user identity Correct answer.
  3. Recomputing preferences from scratch at each session start
  4. Relying on the model's pretrained knowledge of the user

Explanation

Cross-session recall requires persistence outside the ephemeral context, keyed so the correct user's facts can be retrieved later. The context window is discarded between sessions, and pretrained weights contain no user-specific data.

Question 37Agentic AI Concepts and Architecture

An agent must summarize a 500-page contract and then answer precise clause questions about it. Which architecture most appropriately combines the two needs?

  1. A tool that returns only the document's table of contents
  2. An offline summarization plus a retrieval index over clause-level chunks for question answering Correct answer.
  3. Repeated full-document scans on every user question
  4. A single prompt that pastes the entire contract into the context window

Explanation

Precomputing a summary and indexing clause-level chunks lets the agent answer specific questions by retrieving the relevant passages rather than rescanning or truncating the document. This separates broad comprehension from precise lookup, which is the standard pattern for large corpora.

Question 38Agentic AI Concepts and Architecture

Two tools have partially overlapping purposes: one searches internal tickets and one searches public documentation. Which design change most improves the agent's ability to choose correctly?

  1. Merging both tools into a single tool with no parameters
  2. Rewriting each tool's description to state its scope, inputs, and when it should be preferred Correct answer.
  3. Removing descriptions and letting the model infer usage from the tool names
  4. Adding a third tool that performs both searches identically

Explanation

Tool selection depends heavily on descriptions that clearly delineate scope and usage conditions, so disambiguating them is the most direct fix. Removing information or duplicating capabilities makes the choice harder rather than clearer.

Question 39Agentic AI Concepts and Architecture

A multi-agent pipeline passes messages between agents, and a downstream agent sometimes receives malformed context. Which architectural practice most reduces this failure?

  1. Removing the orchestrator so agents communicate peer to peer without checks
  2. Increasing the number of agents in the pipeline
  3. Allowing agents to modify each other's prompts at runtime
  4. Defining a typed message contract between agents and validating payloads at each handoff Correct answer.

Explanation

Typed contracts with validation at each boundary catch malformed or missing fields before they propagate, making failures localized and diagnosable. Free-form prompt mutation and removing coordination increase unpredictability, and adding agents multiplies the handoffs that can break.

Question 40Agentic AI Concepts and Architecture

Which statement best explains why an agent's tool results should be treated as untrusted input when forming the next prompt?

  1. Tool output can contain injected instructions that redirect the agent's behavior Correct answer.
  2. Tool results are never useful for reasoning
  3. Tool results are always shorter than model outputs
  4. Tool output is always encrypted and unreadable

Explanation

External content returned by tools can carry adversarial text that the model may interpret as instructions, a prompt-injection risk. Treating tool output as data rather than instructions, and filtering or delimiting it, prevents hijacking of the agent's control flow.

Question 41Oracle AI Agent Platform and Services

In Oracle's agentic AI architecture, what is the primary role of the agent's reasoning loop when it receives a user request?

  1. It decomposes the goal into steps and selects tools or knowledge sources to satisfy it Correct answer.
  2. It encrypts the payload before forwarding it to the language model
  3. It caches the request so identical prompts bypass the model on later calls
  4. It compiles the request into SQL and executes it directly against the database

Explanation

An agent differs from a plain LLM call because it plans: it interprets the goal, breaks it into sub-tasks, chooses among available tools or retrieval sources, observes results, and iterates until the goal is met. Compilation to SQL, caching, and encryption are supporting infrastructure concerns, not the reasoning loop itself.

Question 42Oracle AI Agent Platform and Services

Which Oracle Cloud Infrastructure service is designed to let developers build, deploy, and manage AI agents that combine LLM reasoning with enterprise data and tools?

  1. OCI Bastion
  2. OCI Object Storage
  3. OCI File Storage
  4. OCI Generative AI Agents Correct answer.

Explanation

OCI Generative AI Agents is the managed service for creating agents that ground responses in enterprise knowledge bases and connect to tools. Object Storage, Bastion, and File Storage are infrastructure services with no agent orchestration capability.

Question 43Oracle AI Agent Platform and Services

A team wants an agent to answer policy questions using internal PDFs without fine-tuning a model. Which capability should they configure?

  1. Few-shot examples embedded in the system prompt
  2. A larger context window on the base model
  3. A retrieval-augmented knowledge base connected to the agent Correct answer.
  4. A higher temperature setting for creative responses

Explanation

Retrieval augmentation lets an agent ground answers in an indexed knowledge base of enterprise documents, avoiding model retraining. Context size, few-shot prompting, and temperature change model behavior but do not supply authoritative internal content at query time.

Question 44Oracle AI Agent Platform and Services

What does a tool definition provide to an Oracle AI agent so the model can invoke it correctly?

  1. A static list of past outputs the model should imitate
  2. A private key used to sign every model response
  3. A name, description, and input schema describing parameters and their types Correct answer.
  4. A compiled binary the agent executes locally on the client

Explanation

Tools are exposed to the model through declarative metadata: an identifier, a natural-language description of purpose, and a parameter schema. The model uses this metadata to decide when and how to call the tool, while the platform handles actual execution.

Question 45Oracle AI Agent Platform and Services

Which statement best describes grounding in the context of Oracle enterprise agents?

  1. Running the model on dedicated hardware to reduce latency
  2. Storing conversation history in a relational table
  3. Restricting the model to a fixed vocabulary of approved words
  4. Anchoring model output to retrieved enterprise content so answers are traceable and current Correct answer.

Explanation

Grounding ties generated responses to retrieved source material, improving factual accuracy and allowing citations back to enterprise content. It is not a vocabulary restriction, a hardware choice, or a persistence mechanism.

Question 46Oracle AI Agent Platform and Services

An agent must call an external REST endpoint to check order status. Which platform capability makes this possible?

  1. Registering the endpoint as a tool the agent can invoke during its reasoning loop Correct answer.
  2. Increasing the model's maximum output tokens
  3. Adding the endpoint URL to the model's system prompt as plain text
  4. Uploading the API documentation to object storage

Explanation

External actions are exposed to agents as tools, which the reasoning loop selects and invokes with structured arguments. Mentioning a URL in a prompt, storing documentation, or raising token limits does not give the agent a callable integration.

Question 47Oracle AI Agent Platform and Services

What is the purpose of an agent session in Oracle's agentic platform?

  1. It maintains conversational state and context across related user interactions Correct answer.
  2. It defines the network firewall rules for the agent endpoint
  3. It compiles agent instructions into container images
  4. It permanently archives all prompts for regulatory audit

Explanation

A session tracks the interaction thread so the agent retains context, prior tool results, and user intent across turns. Auditing, network policy, and image builds are handled by separate platform services.

Question 48Oracle AI Agent Platform and Services

Which benefit most directly justifies using an agent rather than a single prompt-response LLM call for a multi-step procurement task?

  1. The agent guarantees deterministic output for identical inputs
  2. The agent can iteratively gather data, call systems, and adjust its plan based on intermediate results Correct answer.
  3. The agent always produces shorter responses
  4. The agent eliminates the need for any data access controls

Explanation

Agents add value when a task requires multiple dependent steps with intermediate observations that inform the next action. They do not guarantee determinism, remove security requirements, or inherently shorten responses.

Question 49Oracle AI Agent Platform and Services

An agent must query a database and then summarize the rows it retrieves. Which sequence reflects correct agent behavior?

  1. Generate the summary first, then invoke the query tool to confirm it
  2. Invoke every available tool in parallel regardless of relevance
  3. Skip tool invocation and answer from the model's pretrained knowledge
  4. Invoke the query tool, observe the returned rows, then generate a summary grounded in those rows Correct answer.

Explanation

Tool-using agents follow an act-observe-reason pattern: call the tool, incorporate the returned observation, then produce output grounded in that data. Summarizing before retrieval or answering from memory risks unsupported claims, and invoking irrelevant tools wastes resources.

Question 50Oracle AI Agent Platform and Services

A developer wants the agent to answer only questions about HR policy and politely decline unrelated topics. Which configuration most directly enforces this?

  1. A lower top-p sampling value
  2. Instructions that define the agent's scope and refusal behavior Correct answer.
  3. A larger embedding dimension for the knowledge base
  4. A longer session timeout

Explanation

Agent instructions (the system-level guidance) define persona, scope, and how to handle out-of-scope requests. Sampling parameters, embedding size, and session timeouts affect quality or performance but not topical boundaries.