Free Salesforce Certified Agentforce Specialist (AI-201) Sample Questions

Correct answer: A — Use Prompt Builder with a template that references fields.

• A. Correct. Prompt Builder is designed to generate trusted AI outputs from structured Salesforce data using reusable prompt templates that dynamically reference record fields. This enables consistent, context-aware text generation directly within Salesforce.
• B. Incorrect. Manually creating descriptions in a document and uploading them is a static process that does not leverage Prompt Builder's dynamic capabilities or integration with Salesforce data. It does not generate content using AI or field references.
• C. Incorrect. While Apex can concatenate field values to assemble text, it is custom code and not the intended use of Prompt Builder. Prompt Builder is a low-code tool for prompt engineering, not programmatic logic.
• D. Incorrect. Einstein Prediction Builder is used for predictive modeling, such as scoring and classification, not for generating natural-language text. It does not create descriptions or responses like Prompt Builder does.

Correct answer: A — JSON with specific field values

• A. Correct. A Field Generation Template produces structured output, typically in JSON, containing specific field values. This structured format is designed for easy mapping into Salesforce fields or other data systems.
• B. Incorrect. Plain text narrative descriptions are typical of Flex Templates, which allow free-form natural language output, not the structured field-level output of a Field Generation Template.
• C. Incorrect. Although XML is structured, Field Generation Templates in Salesforce use JSON for simplicity and compatibility with the Salesforce data model, not XML.
• D. Incorrect. HTML pages with static content are not the output of Field Generation Templates, which focus on generating dynamic, structured field values rather than full web pages.

Correct answer: B — To ensure the response can be used by other systems.

• A. Incorrect. While formatted output may sometimes reduce tokens, this is not the primary reason. The main goal is to make the output predictable and usable, not to optimize efficiency.
• B. Correct. Specifying the output format ensures the response is structured predictably, making it easy to parse by downstream systems, APIs, or automation workflows. This consistency is essential for integration.
• C. Incorrect. Adding format instructions often makes the prompt longer. The primary purpose is not to shorten the prompt but to control the structure of the output for better usability.
• D. Incorrect. Format instructions are designed to reduce randomness and enforce consistency, not to introduce variety. Standardized outputs are needed for reliability and interoperability.

Correct answer: A — It defines the object that provides the record context for the prompt.

• A. Correct. The Base Object field defines the primary Salesforce object (e.g., Account, Case, Lead, Opportunity) that provides the record context for the prompt template. This allows the prompt to access and reference fields from a specific record during execution, enabling personalized responses.
• B. Incorrect. User permissions and who can execute a prompt template are managed through profiles, permission sets, sharing settings, or the execution context (e.g., where the template is invoked), not by the Base Object field.
• C. Incorrect. The output format of the generated response is determined by the prompt template type (e.g., free text, structured) and the model's configuration, not by the Base Object field. The Base Object only establishes the record context.
• D. Incorrect. While a Flow can be used to trigger automation, the Base Object field does not select or configure the Flow that triggers prompt execution. The Flow is selected separately in the prompt template setup or automation configuration.

Correct answer: A — Add a Prompt Template action to the dialog in Einstein Bot Builder.

• A. The admin adds a Prompt Template action to the dialog in Einstein Bot Builder, which allows the bot to reference and execute the template at runtime using the current conversation context, such as the chat transcript. This is the standard method to integrate prompt templates into bot flows.
• B. Invoking via an Apex class action is not required; Salesforce provides a native Prompt Template action designed specifically for this purpose. Prompt templates are configured directly in the bot dialog, not through custom Apex.
• C. Prompt templates are not assigned to user profiles. Access is handled through bot configuration and the template/action setup, not via individual user permissions.
• D. There is no process to import a prompt template as a bot skill. Bot skills are separate entities; prompt templates are directly added as actions within the bot dialog.

Correct answer: C — Prompt Chaining

• A. Incorrect. Knowledge grounding uses a predefined knowledge base to inform responses, but it does not inherently support sequential multi-step reasoning. The scenario requires a flow where each step depends on the previous output, which knowledge grounding alone cannot orchestrate.
• B. Incorrect. Data grounding involves grounding responses in external or enterprise data, but it is not designed for chaining multiple dependent prompts. While data may be used, the core need is to execute steps sequentially, not just to reference data.
• C. Correct. Prompt chaining breaks a complex task into multiple sequential prompts, where the output of one prompt becomes the input for the next. This technique perfectly fits the described workflow: first generating a summary, then using that summary to retrieve a solution from a knowledge base.
• D. Incorrect. Model fine-tuning modifies the model's behavior through training on specific examples, which is unnecessary for a runtime multi-step workflow. The scenario is better addressed with prompt orchestration rather than retraining the model.

Correct answer: B — Apply Knowledge Grounding to products, Data Grounding to pricing and accounts.

• A. Incorrect. Knowledge Grounding is best suited for unstructured or semi-structured content like product documentation. Pricing and account data are structured and not typically stored in a knowledge base, making a combined Knowledge Grounding source inadequate.
• B. Correct. Knowledge Grounding is appropriate for product details in the knowledge base (unstructured/semi-structured), while Data Grounding is designed for structured data such as price books and Account records. This combination matches each data type to the optimal grounding method.
• C. Incorrect. A manually scripted Prompt Chain can retrieve datasets sequentially but is more complex and less efficient than using built-in grounding techniques, which are optimized for these use cases.
• D. Incorrect. Data Grounding is ideal for structured CRM data but cannot effectively retrieve unstructured content from a knowledge base. Relying solely on Data Grounding would miss product details.

Correct answers: A, D — It enables retrieval of more granular and relevant information.; It allows the retriever to focus on specific sections rather than whole documents.

• A. Correct. Chunking breaks large documents into smaller, manageable pieces, enabling retrieval of more granular and relevant information during searches. This is a core benefit for retrieval-augmented generation.
• B. Incorrect. Chunking does not primarily reduce computational cost of generating embeddings; embeddings are still created for each chunk. The main benefit is improving retrieval relevance and precision, not lowering embedding cost.
• C. Incorrect. Chunking does not directly improve natural language generation quality of the agent. It enhances retrieval accuracy, which may indirectly support better responses, but NLG quality depends on the LLM, prompting, and context.
• D. Correct. By splitting documents into sections, the retriever can target specific chunks instead of scanning entire documents. This improves precision and efficiency, allowing focus on relevant sections.

Correct answer: B — An inverted index that maps terms to chunks containing those terms

• A. Incorrect. ANN search is used for dense vector similarity, not for sparse indexing. Sparse indexing relies on term matching rather than dense vector embeddings.
• B. Correct. A sparse index typically uses an inverted index structure that maps terms to the chunks or documents containing them, enabling efficient keyword-based retrieval.
• C. Incorrect. A decision tree is not a standard structure for sparse indexing. Hierarchical partitioning is more associated with certain specialized approaches, not general text indexing.
• D. Incorrect. A forward index stores chunks by content or order, but does not support efficient term-to-chunk lookup. That role is fulfilled by an inverted index.

Correct answer: B — Schedule a nightly data refresh in Data Library

• A. Incorrect. Apex triggers on Knowledge__kav are not a supported or reliable method for reindexing the Data Library. The Data Library index is managed separately from standard Salesforce triggers, and reindexing is typically handled by platform-supported scheduled refresh capabilities rather than custom record-trigger logic.
• B. Correct. Scheduling a daily data refresh in the Data Library collection settings is the recommended and most reliable approach because it uses the built-in, supported mechanism for keeping the index synchronized on a recurring basis. This ensures the index is rebuilt automatically on a set schedule without requiring manual intervention or custom code.
• C. Incorrect. Manually clicking 'Refresh Index' each morning is not reliable because it depends on human action and does not guarantee the index is rebuilt every night. It also does not satisfy the requirement for an automated nightly process and is prone to errors or omissions.
• D. Incorrect. The Einstein API is not the standard method for triggering a Data Library index rebuild, and it does not provide functionality for this purpose. This introduces unnecessary complexity compared with native scheduled refresh options. Using supported Data Library settings is more dependable and maintainable.

Correct answers: A, B — The document file size and indexing limits; Whether the document uses an unsupported binary format

• A. Correct. The file size of each document and overall indexing limits are critical constraints. Agentforce has maximum file size and document count limits for collections; exceeding them can prevent successful indexing.
• B. Correct. If a document is in a proprietary binary format that Agentforce cannot extract text from, indexing will fail. Supported formats are required for text extraction.
• C. Incorrect. The number of revisions does not affect indexing. Only the current version of the document matters for content and format.
• D. Incorrect. The author's department is metadata that does not impact the technical ability to index. Indexing depends on content and format, not organizational metadata.
• E. Incorrect. Visual elements like cover color are irrelevant to text extraction. Indexing processes textual content, not visual styling.

Correct answers: C, D — Route to different nodes based on a condition.; Evaluate a boolean and take one of two paths.

• A. Incorrect. Branch nodes do not present static menus; that functionality is handled by a menu node or similar UI component.
• B. Incorrect. Branch nodes cannot directly invoke external APIs; API calls are managed by dedicated action or integration nodes.
• C. Correct. A branch node routes execution to different subsequent nodes based on conditions, enabling dynamic control flow in the agent script.
• D. Correct. A branch node evaluates a boolean expression and follows either a true or false path, implementing conditional logic in the flow.
• E. Incorrect. Dynamic AI responses are generated by AI or generative nodes, not by branch nodes, which are limited to decision-making.
• F. Incorrect. Immediate handoff to a human agent is handled by a dedicated handoff or escalation node; branch nodes only decide routing, not execute transfers.

Correct answer: B — Collapsible code blocks

• A. Incorrect. Inline prompts are used to embed guidance or instructions within the script, but they do not control whether branches can be collapsed. They aid readability and editing context, not structural folding.
• B. Correct. Collapsible code blocks are the Script View feature that lets users expand or collapse branches, simplifying navigation and reducing visual clutter. This is especially useful in larger scripts with nested logic.
• C. Incorrect. Auto-layout helps organize the visual arrangement of script elements, but it does not enable collapsing branches. It primarily arranges nodes in Canvas View, not Script View.
• D. Incorrect. Syntax highlighting improves code readability by color-coding script elements, but it does not provide branch collapsing functionality. It is a formatting aid, not a structural navigation feature.

Correct answer: D — Topic

• A. An Action performs a task or invokes a capability, but it does not decide which conversational path to follow. Actions are executed after the agent has identified the relevant flow, so they do not handle input matching.
• B. An Instruction provides guidance on how the agent should behave or respond, but it does not map user input to a specific conversation path. The routing decision is handled by topics, not instructions alone.
• C. A Knowledge Base contains information the agent can reference to answer questions, but it does not route the conversation to a path. It supports grounding responses, not matching intent to a flow.
• D. A Topic matches user input to a specific conversation path based on intent or subject matter. Once matched, the agent follows the topic's associated flow, instructions, and actions, allowing for the appropriate conversational path.

Correct answer: A — Create a Topic with phrases for manager requests and a transfer Action.

• A. Correct. Topics in Agentforce are designed to detect specific customer intents or phrases (like 'I want to speak to a manager') and can trigger associated Actions, such as transferring to a human agent. This is the appropriate place to implement a rule that responds to a specific phrase with an immediate transfer.
• B. Incorrect. While Agent Instructions define behavioral rules and guidelines for the agent, they are not the primary mechanism for detecting specific phrases and executing immediate actions like transfers. Instructions are better suited for general tone and behavior, not for conditional phrase-triggered actions.
• C. Incorrect. An Escalation Policy is used to define broader conditions (e.g., unresolved cases, sentiment analysis) for routing or escalating, but it is not designed for specific phrase-based triggers like a direct request for a manager. Such a rule is better handled by a Topic.
• D. Incorrect. Knowledge Articles provide informational content to customers or the agent, but they do not enforce transfer logic or trigger actions. They can advise on how to request a manager, but they cannot automatically execute the transfer.

Correct answers: A, B — Send a session ID with each request to maintain conversation state.; Use OAuth 2.0 for secure authentication and authorization when calling the Agent API.

• A. Correct. The Agent API requires session management to maintain conversation context across multiple requests. Sending a session ID with each request preserves state and is considered a best practice for multi-turn interactions.
• B. Correct. OAuth 2.0 is the recommended authentication and authorization method for Salesforce APIs, including the Agent API. It provides secure, scoped access and avoids exposing long-lived secrets.
• C. Incorrect. Storing user credentials locally in a mobile app is insecure and violates security best practices. Credentials can be exposed if the device is compromised. Use OAuth tokens instead.
• D. Incorrect. Agent interactions typically require state management across multiple requests. A single API call without session management is not feasible for maintaining conversation continuity.
• E. Incorrect. Embedding an API key directly in source code is insecure, as the key can be extracted from the app. API keys should be securely managed, not hardcoded.

Correct answers: A, B — The need for a fully customized user interface that matches the portal's design.; The requirement to integrate the agent with other non-Salesforce systems.

• A. Correct. Agent API is appropriate when you need to embed agent capabilities into a fully customized user interface, such as a portal with its own design and interaction patterns. It provides programmatic flexibility to build and orchestrate the agent behavior outside of standard Salesforce UIs.
• B. Correct. Agent API is well-suited for integrating agent functionality with external or non-Salesforce systems because it can be called from custom applications and services, enabling cross-platform agent interactions.
• C. Incorrect. Agent API requires code or scripting to interact with agents. If the goal is to avoid code and use only declarative tools, alternatives like Flows or Agentforce Builder are more appropriate.
• D. Incorrect. Standard Salesforce Lightning pages work with agents using declarative tools or out-of-the-box components. Agent API is unnecessary for this scenario, as it is designed for custom programmatic access, not standard page interaction.
• E. Incorrect. One-time batch analysis of historical case data is a reporting or analytics task, best handled by tools like Salesforce Reports, Dashboards, or Einstein Analytics. Agent API is meant for interactive agent integrations, not offline batch processing.

Correct answer: A — Use a template expression in the email text to insert agent variables like customer name and order number.

• A. Correct. Template expressions in Agentforce allow you to dynamically insert agent variables (e.g., `{!customerName}`, `{!orderNumber}`) directly into email text, ensuring real-time data insertion and consistent content.
• B. Incorrect. While Flows can concatenate strings, they are not the primary or recommended method for dynamically inserting agent variables into emails in Agentforce. Template expressions are more efficient for this use case.
• C. Incorrect. Formula fields on records are static and cannot dynamically reference agent variables at runtime. They are not suitable for real-time email personalization in Agentforce.
• D. Incorrect. Custom metadata types are useful for storing configuration data, but they are not the mechanism for token replacement at send time in an agent email action. Template expressions are the correct approach.

Correct answer: A — Add a filter on the follow-up email that checks if the case status variable is not equal to 'Closed'.

• A. Correct. Adding a filter on the follow-up email action that checks the case status variable is a deterministic mechanism. It programmatically prevents the action from executing when the condition is false, ensuring the agent never sends a follow-up on closed cases without relying on interpretation.
• B. Incorrect. A prompt reminder is not deterministic; it relies on the agent's LLM interpretation of the instruction, which may not consistently block the action. Filter-based or variable-based controls are preferred for deterministic behavior.
• C. Incorrect. Manually disabling the email action is not scalable or automated. It requires human intervention each time a case closes, which is not a built-in agent mechanism like filters or variables.
• D. Incorrect. A blank template does not prevent the email action from being executed. It may still send an empty email, and it is not a deterministic filter-based or variable-based control to block the action.

Correct answer: B — Copy the standard topic and then modify the duplicate to add the required action.

• A. Incorrect. Standard topics in Agentforce are predefined and locked, so they cannot be directly edited in Agentforce Builder to add actions from the library. This is not a supported method.
• B. Correct. Standard topics cannot be directly extended. The supported approach is to copy the standard topic, which creates a custom topic, and then modify the duplicate to add the required action. This preserves the original standard topic while allowing customization.
• C. Incorrect. Standard topics are not defined by an Apex class that administrators can directly edit in Salesforce Setup. This is not the intended mechanism for modifying Agentforce topics.
• D. Incorrect. While creating a custom action and using it in a new custom topic is a valid approach, it does not address the requirement of including an additional action in an existing standard topic. The question asks specifically about modifying a standard topic.

Correct answer: A — The 'View Order' standard action is missing from the Order Management topic.

• A. Correct. The 'View Order' standard action is required for the agent to retrieve order details and respond with status updates. Without this action configured on the Order Management topic, the agent can recognize the intent but cannot fulfill the request, even though the topic is active and other intents work.
• B. Incorrect. Field-level security (FLS) affects data visibility but does not prevent the agent from recognizing or processing the intent. The agent would still attempt to respond, and the issue described is a topic configuration problem, not a permissions issue.
• C. Incorrect. Since the topic is active and the agent understands other intents, the customer utterances are likely matching the topic's defined examples. The failure is in fulfillment, not intent recognition, so missing utterances are not the cause.
• D. Incorrect. Standard topic behavior does not depend on Einstein training data; it relies on built-in actions and configured topics. The agent's ability to understand other intents indicates that the model is working, so a missing action is the more likely issue.

Correct answer: B — The Pre-Chat Page Layout in Salesforce Setup

• A. Incorrect. The Chat Button in Embedded Service setup controls how the chat launcher appears and behaves, but it does not define the pre-chat form fields. Pre-chat data collection is configured separately from the button.
• B. Correct. The Pre-Chat Page Layout in Salesforce Setup is used to define the fields and layout shown to the user before a chat starts. This component allows administrators to collect contact details ahead of the conversation, and it applies to Agentforce chat on Experience Cloud sites.
• C. Incorrect. The Community Builder component for Chat is used to place or manage the chat widget on an Experience Cloud site, but it does not configure the pre-chat form fields. The form configuration is handled in the chat setup, not the page builder.
• D. Incorrect. The Live Agent Configuration in Omni-Channel handles routing and agent availability, not the pre-chat form design. It helps manage chat delivery but does not define the pre-chat data entry experience.

Correct answer: A — The Omni-Channel Supervisor Dashboard

• A. Correct. The Omni-Channel Supervisor Dashboard natively provides real-time and historical metrics for agents across channels, including chat and email. It allows administrators to view workload distribution and performance breakdowns by channel, making it the appropriate tool for comparing chats versus emails per agent.
• B. Incorrect. While the Agentforce Analytics app from AppExchange can offer additional analytics, it is not the standard, built-in tool for accessing channel-specific metrics in Salesforce. The primary location for such reports is the Omni-Channel Supervisor Dashboard.
• C. Incorrect. Standard Salesforce Reports on Agent Activity are too generic and do not natively provide the channel-level breakdown needed to compare chats versus emails handled by each agent without significant customization.
• D. Incorrect. The Service Console Performance tab in Setup is used for monitoring the performance and configuration of the Service Console itself, not for tracking agent-level channel metrics like chat and email counts.

Correct answer: B — Review failed test case transcripts to identify gaps and update knowledge articles or topics.

• A. Incorrect. Adding more test cases can improve coverage, but it does not directly address why the current test cases are failing. The immediate next step should be to analyze the failures and understand the root cause.
• B. Correct. Reviewing failed test case transcripts helps identify where the agent misunderstood the request, lacked knowledge, or chose the wrong action. Based on those insights, you can update knowledge articles, topics, or other configuration to improve performance.
• C. Incorrect. Re-running tests can help confirm whether a failure is consistent, but it does not solve the underlying issue. It is useful as a diagnostic check, not the primary corrective action after failures are observed.
• D. Incorrect. Changing expected outcomes to match the agent's current behavior may hide real defects instead of fixing them. Test cases should validate desired behavior, so expectations should only be updated if the intended behavior truly changed.

Correct answer: B — Recall, which measures completeness of predictions.

• A. Incorrect. Precision measures the proportion of positive identifications that were actually correct (true positives / (true positives + false positives)). It focuses on the exactness of positive predictions, not the proportion of actual positives identified.
• B. Correct. Recall (also called sensitivity or true positive rate) measures the proportion of actual positives that were correctly identified by the agent (true positives / (true positives + false negatives)). It reflects the completeness of capturing relevant cases.
• C. Incorrect. The F1 score is the harmonic mean of precision and recall, balancing both metrics, but it does not directly measure the proportion of actual positives identified.
• D. Incorrect. Accuracy measures the overall proportion of correct predictions across all classes (correct predictions / total predictions), not specifically the identification of actual positives. It can be misleading when classes are imbalanced.

Correct answers: A, B, E — Single-turn evaluation; Multi-turn evaluation; Batch evaluation

• A. Single-turn evaluation is a valid Testing Center evaluation type used to test individual interactions against expected outputs.
• B. Multi-turn evaluation is a valid Testing Center evaluation type used to test conversational flows across multiple exchanges, assessing context retention and appropriate responses.
• C. A/B testing evaluation is not a standard Testing Center evaluation type; it is a broader experimentation pattern typically used for comparing features or models.
• D. Real-time monitoring evaluation is not a Testing Center evaluation type; it belongs to operational observability and deployment monitoring.
• E. Batch evaluation is a valid Testing Center evaluation type used to assess multiple test cases at scale efficiently.
• F. Shadow deployment evaluation is not a Testing Center evaluation type; it refers to a deployment strategy, not an evaluation method.

Correct answer: A — Change sets

• A. Correct. Change sets require a deployment connection between the source (sandbox) and target (production) orgs, established via the Salesforce UI, relying on a trust relationship.
• B. Incorrect. The Metadata API does not inherently require a direct org-to-org connection; it can be used via tools like Workbench, CLI, or CI/CD pipelines that authenticate to each org separately.
• C. Incorrect. Unmanaged packages are uploaded to the source org and installed in the target org via a URL or package ID, without requiring a direct connection between the two orgs.
• D. Incorrect. Managed packages are distributed via AppExchange or a direct install link, and do not require a direct connection between sandbox and production orgs.

Correct answer: A — Change sets provide a declarative interface for selecting components to deploy.

• A. Correct. Change sets provide a declarative, point-and-click interface within the Salesforce UI, allowing administrators to select and deploy components without writing code or using external tools. This makes change sets ideal for simple configuration changes like Agentforce templates, whereas the Metadata API requires scripting and is more technical.
• B. Incorrect. Change sets have file size limitations (e.g., 39 MB for most components) that are typically smaller than what the Metadata API can handle. The Metadata API is more suitable for larger deployments, and change sets are not designed for larger file sizes.
• C. Incorrect. Change sets are designed for deployment between two connected orgs and can only deploy to a single target org at a time. The Metadata API, when used with tools like Salesforce CLI or CI/CD pipelines, can deploy to multiple orgs simultaneously.
• D. Incorrect. Change sets do not automatically resolve all dependencies. While Salesforce may identify some related components, administrators must manually review and include all required dependencies (e.g., custom fields, objects) in the change set to avoid deployment errors.

Correct answer: A — Agent Performance Dashboard

• A. Correct. The Agent Performance Dashboard provides aggregated metrics and insights into agent utilization, conversation outcomes, and performance trends over defined time periods, including fiscal quarters. It is designed for high-level monitoring and governance.
• B. Incorrect. Agent Builder analytics focus on the development, configuration, and deployment of agents, not on tracking utilization or conversation outcomes over time.
• C. Incorrect. Conversation Transcript Logs store individual conversation details for review and troubleshooting, but they do not aggregate or track utilization or outcomes across a fiscal quarter in a dashboard format.
• D. Incorrect. Field Service Settings are used for configuring field service operations and are unrelated to agent performance tracking or analytics.

Correct answer: A — Agent Analytics in the Analytics Studio with pre-built dashboards.

• A. Correct. Agent Analytics in Analytics Studio provides pre-built dashboards specifically designed to analyze agent performance metrics, including peak usage times and conversation duration patterns. It offers out-of-the-box visualizations for trend analysis without requiring custom builds.
• B. Incorrect. While a custom Report Type on the Conversation object can provide basic data, it lacks the pre-built analytics and visualizations needed for quick, comprehensive analysis of usage patterns. Agent Analytics is the more appropriate tool for this purpose.
• C. Incorrect. Exporting Agent Logs to an external BI tool requires additional setup, data movement, and custom modeling. The built-in Agent Analytics offers immediate, pre-built insights for analyzing usage trends and conversation durations.
• D. Incorrect. The real-time monitor on the Agentforce Console home page is useful for live operational visibility, but it does not provide historical trend analysis needed to identify peak usage times and conversation duration patterns.

Correct answer: B — Analyze dialog flows to find inefficiencies and optimize paths

• A. Reverting to a previous version may stop the issue temporarily but does not address the root cause of poor performance. Agent optimization requires understanding the current design's friction points rather than rolling back without analysis.
• B. Analyzing dialog flows helps identify where users get stuck or where unnecessary steps exist. Optimizing conversation paths directly improves agent efficiency and overall performance, making this the most effective action.
• C. Increasing the confidence threshold may reduce low-confidence actions but does not inherently optimize the conversation path or improve dialog design. It can also increase 'I don't know' responses and reduce responsiveness, negatively impacting user experience.
• D. Disabling the agent during peak hours prevents observation under real demand and does not help optimize behavior. The goal is to analyze performance and refine the experience, not to avoid using the agent during busy times.

Correct answers: A, B, C, E — Retrain the NLU model using recent conversation logs; Increase the confidence threshold to reduce misclassification; Review user feedback to identify common pain points; Adjust dialog flows to be more empathetic and accurate

• A. Retraining the NLU model with recent conversation logs helps the agent adapt to new language patterns and improve intent recognition, a standard optimization based on analytics.
• B. Increasing the confidence threshold reduces misclassification by acting only on high-confidence predictions, improving accuracy. However, it should be balanced to avoid excessive fallbacks.
• C. Reviewing user feedback is a key practice to identify recurring issues and pain points, enabling targeted improvements to intents, dialogs, and overall agent performance.
• D. Simply adding synonyms without adjusting dialog flows does not address underlying conversation logic or user experience issues. It is an incomplete optimization.
• E. Adjusting dialog flows to be more empathetic and accurate directly improves user satisfaction and resolution rates, making it a valid optimization based on feedback and analytics.
• F. Reducing the number of intents can oversimplify the agent's capabilities, potentially harming performance on diverse queries. It is not generally recommended as a primary optimization.

Correct answer: B — SOMA’s single-agent approach cannot scale across diverse, high-volume business functions.

• A. Incorrect. While a single-agent design reduces coordination overhead, consolidation does not inherently enable easy scaling. The centralized nature becomes a bottleneck as task diversity and volume increase, limiting scalability despite the reduction in overhead.
• B. Correct. A SOMA architecture is generally unsuitable for diverse, high-volume business functions because the single agent cannot distribute work across specialized agents or services. Its centralized design becomes a performance constraint and scalability bottleneck.
• C. Incorrect. SOMA does offer strict control and consistent outputs, but this does not make it the best option for scalability. In scenarios requiring broad scale and varied functions, the single-agent approach is likely to struggle, and other architectures may be more appropriate.
• D. Incorrect. Deploying SOMA on powerful infrastructure can improve performance, but it does not resolve the fundamental architectural limitation of a single point of execution. Scaling by brute force still leaves the system vulnerable to bottlenecks and does not address the need for specialization in high-volume, diverse tasks.

Correct answer: B — A2A facilitates direct messaging for capability exchange during initial handshake.

• A. Incorrect. MCP (Model Context Protocol) does not maintain a distributed directory of agent functions. It is a client-server protocol for connecting models to tools and context sources, not a directory service for agent skills.
• B. Correct. A2A (Agent-to-Agent) protocol is designed for direct communication between agents, including exchanging capability details during the handshake process. This enables agents to discover whether another agent can perform a required task.
• C. Incorrect. While both protocols support discovery, the description of roles is inaccurate. MCP discovers tools/resources but does not act as a registration directory, and A2A focuses on capability exchange, not authentication management.
• D. Incorrect. Discovery is within the scope of both protocols. MCP includes tool and resource discovery, and A2A supports capability discovery between agents, so a separate orchestration layer is not strictly required for discovery.

Correct answers: A, B, C — Enables agents from different vendors to interoperate without custom integrations.; Improves scalability by supporting decentralized agent communication and coordination.; Reduces integration effort by leveraging standardized messaging and discovery mechanisms.

• A. Correct. Open standard multi-agent protocols like MCP and A2A are designed to enable interoperability between agents from different vendors, eliminating the need for custom point-to-point integrations. This is a core goal of such standards.
• B. Correct. These protocols support decentralized communication and coordination patterns, which improves scalability by allowing agents to operate independently and coordinate directly without relying on a single central orchestrator.
• C. Correct. Standardized messaging, tool access, and discovery mechanisms reduce the engineering effort needed to connect agents and services. Using a common protocol avoids repeated custom integration work.
• D. Incorrect. While protocols may aim for efficiency, they do not inherently guarantee low-latency responses or enforce quality of service at the protocol level. Performance depends on implementation, network, and infrastructure.
• E. Incorrect. Open standard protocols do not automatically ensure adherence to global data residency and privacy regulations. Compliance is typically handled at the implementation or governance layer, not by the protocol itself.
• F. Incorrect. Built-in negotiation templates for complex business transactions are not a standard feature of MCP or A2A. Such workflows are usually implemented by the application or orchestration layer.