Explanation:

When using the Agentforce Testing Center to evaluate agent responses, standard metrics can tell you if an agent followed simple instructions or if the text is structured correctly. However, analyzing highly specific qualitative traits—like a company's unique tone, brand voice guidelines, or complex internal quality rubrics—requires customized grading logic.

Why C is correct:
Standard, out-of-the-box evaluations are rigid and do not know your corporate brand identity. Agentforce allows specialists to define Custom Evaluations powered by an LLM Judge. By configuring a custom LLM judge, you can supply your company's precise brand-voice playbook and criteria as prompt parameters. The testing engine will then use that specialized context to grade your agent's test-run outputs on a nuanced scale rather than a binary pass/fail check.

Why A is incorrect:
The Coherence evaluation is a standard metric that looks for logical flow, clarity, and grammatical consistency. It measures whether a response makes general sense, but it cannot evaluate whether the text aligns with Universal Containers' specific brand voice.

Why B is incorrect:
The default Response Evaluation metrics cover foundational traits like helpfulness, accuracy, and safety. They are not tailored to custom corporate guidelines.

Reference
Salesforce Help / Agentforce Developer Guide: Testing Center Custom Metrics. For advanced quality assurance that goes beyond standard evaluation parameters, administrators can create custom metrics utilizing a LLM judge. This allows organizations to provide custom guidelines, scoring definitions, and rubrics to monitor brand alignment and response voice.

Explanation:

The key distinction is who decides whether the action executes.

reasoning.actions
Actions listed here are available to the LLM as tools.
The model chooses whether or not to invoke them based on its reasoning and the current conversation.
There is no guarantee that a listed action will be executed on every applicable turn.

run @actions.
This is an explicit instruction to execute the specified action.
It bypasses the LLM's discretionary tool selection and ensures the action runs deterministically whenever that line is reached in the Agent Script.
This approach is appropriate for mandatory business logic, such as:
- Checking cancellation eligibility
- Performing compliance or security validations
- Verifying customer identity
- Logging audit events

Since the requirement is to guarantee that "Check Cancellation Eligibility" executes on every relevant turn, using run @actions.CheckCancellationEligibility is the correct implementation.

Why the Other Options Are Incorrect

A. ❌ Incorrect
reasoning.actions does not automatically execute actions on every turn. It merely exposes them as tools the LLM may use. Therefore, the difference is not merely syntactic.

B. ❌ Incorrect
The run command is not limited to being nested inside reasoning.actions, nor is reasoning.actions required for the LLM to access an explicitly invoked action. run is used specifically to force execution, not simply to pass parameters.

Reference
- Salesforce Agentforce documentation on Agent Scripts and action invocation (concepts of deterministic execution vs. LLM-selected tools).
- Salesforce Agentforce Specialist (AI-201) Exam Guide covering Agent Scripts, reasoning, and action orchestration.

Exam Tip
A common exam pattern is distinguishing deterministic execution from LLM-driven reasoning:

- reasoning.actions → The LLM decides whether to call the action.
- run @actions. → The action always executes when that step is reached.

If a question includes phrases such as "must always execute," "guaranteed," "mandatory," "deterministic,"or "without relying on the reasoning engine,"the correct choice is typically the explicit run invocation.

Explanation:

Why B is correct:
When creating a prompt template that needs to consistently extract specific structured data (like product model numbers and quantities) from unstructured text (an email) and then draft a response, precision and reliability are the top priorities.

Clear, positive instructions tell the LLM exactly what to do (e.g., "Extract the product model number and quantity from the following email. Format your response as a draft reply to the customer."). Positive instructions (telling it what to do rather than what not to do) reduce confusion and improve output quality.

Few-shot examples (providing 1–3 examples of input emails and the exact desired output format) are the single most effective technique for guiding an LLM to produce consistent, formatted outputs. Few-shot learning anchors the model to the specific schema, format, and extraction logic you expect, dramatically improving reliability for structured data extraction tasks.

Why A is incorrect:
Open-ended questions are useful for creative brainstorming or exploratory conversations, but they are counterproductive for structured data extraction. Asking open-ended questions (e.g., "What can you tell me about the product?") will result in verbose, unpredictable, and inconsistent outputs that do not reliably capture the specific model number and quantity fields. This approach introduces variability, exactly what you want to avoid for a template intended for consistency.

Why C is incorrect:
The temperature setting controls the randomness/creativity of the LLM's output. A high temperature increases randomness and creativity, which is useful for creative writing but disastrous for structured extraction tasks. High temperature would cause the model to hallucinate, vary its output format, and extract inconsistent information. For extraction and drafting responses, you want a low temperature (closer to 0) to ensure deterministic, repeatable, and accurate outputs—not high temperature.

References:
Salesforce Official Documentation: "Prompt Engineering Best Practices for Agentforce" – recommends using clear, explicit instructions combined with few-shot examples for structured extraction tasks to maximize consistency and accuracy.

Explanation:

When configuring a file upload-based Agentforce Data Library, Salesforce enforces different size limits depending on file type: files must stay within size limits of 4 MB for text/HTML and 100 MB for PDFs.

This distinction makes practical sense:
Text/HTML files are plain text-based and typically much smaller in raw size, so a 4 MB cap comfortably covers large documents while keeping processing fast.
PDF files often contain embedded formatting, images, and layout data that inflate file size well beyond their actual text content, so a much larger 100 MB cap is needed to accommodate real-world business documents (manuals, policy documents, catalogs, etc.) while Data Cloud extracts and chunks the underlying text content.

Why not A:
This reverses the limits — it's not 100 MB for text/HTML and 4 MB for PDFs. Text/HTML has the smaller limit (4 MB), and PDF has the larger limit (100 MB).

Why not B:
The limits are not the same for both file types (50 MB each). Salesforce applies distinct limits per file type as noted above.

Reference:
Salesforce Trailhead – Create a Data Library for AI Agents / Build a File-Based Data Library for an Agent: Ensure the files are within the size limits (4 MB for text/HTML, 100 MB for PDFs).

Explanation:

Option A (Correct):
The Retriever Playground is designed for no-code testing of retrievers. It allows specialists to input queries and directly observe which chunks are surfaced, making it ideal for validating relevance before production deployment.

Option B:
The Agentforce Testing Center is focused on regression testing of agents and subagents, measuring metrics like Subagent Pass % and Action Pass %. It does not provide direct visibility into retriever chunk relevance.

Option C:
The Data 360 Query Editor is used for querying structured data in Data Cloud. It is not intended for testing retrievers or evaluating semantic chunk retrieval.

Reference:
Salesforce documentation on Retriever Playground highlights its role as the declarative, no-code environment for testing retrievers and validating chunk relevance prior to production rollout.