Agentforce is Salesforce's suite of autonomous AI agents built on the Einstein 1 Platform, designed to help businesses scale their operations through intelligent automation. For a Platform Administrator, understanding Agentforce capabilities is essential for deploying and managing these AI-powered assistants across an organization.

Core Capabilities: Agentforce agents can reason, plan, and take actions using large language models combined with company data stored in Salesforce. Administrators configure agents through the Agent Builder, defining topics, actions, and guardrails. Agents can access CRM data, Knowledge articles, and Flows to complete tasks autonomously. The Atlas Reasoning Engine allows agents to interpret user requests, retrieve relevant information, and execute multi-step processes while respecting security and sharing settings.

Key business use cases span multiple departments. In customer service, Agentforce Service Agents resolve cases around the clock, answering questions and processing routine requests, which frees human agents for complex issues. In sales, Sales Development Representative agents can nurture leads, schedule meetings, and answer prospect questions, accelerating the pipeline. Marketing teams use agents to generate campaign content and personalize customer journeys. Commerce agents assist shoppers with product recommendations and order tracking.

Administrators play a critical role in ensuring agents operate safely. This includes setting up permission sets, defining data access through sharing rules, configuring topics that scope what an agent handles, and establishing escalation paths to human employees when needed. Testing agents in the Agent Builder sandbox before deployment helps validate accurate behavior.

The business value comes from increased efficiency, faster response times, reduced operational costs, and improved customer satisfaction. Agents work alongside employees rather than replacing them, handling repetitive tasks so teams focus on high-value work.

For certification, focus on how agents integrate with Data Cloud, Flows, and Apex actions, plus governance practices such as monitoring agent performance through analytics dashboards and refining instructions based on real interactions to continuously improve outcomes across the organization.

In Agentforce, selecting the correct agent type is a foundational decision that shapes how your agent behaves, who interacts with it, and what data it can access. Salesforce offers two primary categories: customer-facing agents and employee-facing agents. Understanding the distinction helps administrators deploy agents that align with business goals.

Customer-facing agents are designed to interact with external users such as customers, prospects, or partners. A common example is Agentforce Service Agent, which handles support inquiries on websites, messaging channels, or communities. These agents operate with carefully scoped permissions to ensure that only appropriate, public-safe information is shared with external audiences. Security and guardrails are critical here, since these agents engage people outside your organization. They typically resolve cases, answer product questions, and escalate to human agents when needed.

Employee-facing agents serve internal users like sales representatives, service teams, or staff members. Agentforce Sales and internal productivity agents fall into this group. Because employees already have authenticated access to Salesforce, these agents can leverage broader internal data, surface CRM records, draft emails, summarize accounts, and assist with day-to-day tasks. The trust boundary differs, allowing richer access to sensitive company information that would not be suitable for external sharing.

When choosing between the two, administrators should evaluate the intended audience, the sensitivity of accessible data, required channels, and the desired outcomes. Consider whether the interaction happens inside authenticated Salesforce experiences or on public channels. Also weigh compliance requirements, since customer-facing scenarios demand stricter controls.

Proper agent type selection ensures that topics, actions, and knowledge sources match the user population. It also streamlines configuration of permissions, connections, and guardrails. By matching the agent type to the real business need, administrators create secure, effective experiences that boost productivity for employees while delivering helpful, trustworthy service to customers across every supported channel.

In Agentforce, AI agents operate through a structured framework built on four key components: Topics, Instructions, Actions, and Reasoning. Understanding how these elements interact helps administrators build effective, autonomous agents on the Salesforce Platform.

Topics define the scope of what an agent can handle. A topic represents a specific job or area of expertise, such as answering billing questions or processing order returns. Each topic groups related capabilities together, allowing the agent to understand the boundaries of its responsibilities and route requests appropriately.

Instructions provide the guidelines that shape how an agent behaves within a topic. These are natural language statements that tell the agent what it should do, how it should respond, and any rules it must follow. Well-written instructions ensure the agent stays aligned with business policies and delivers consistent, accurate results while respecting company guardrails.

Actions are the concrete tasks an agent can perform to fulfill a user's request. Actions might include querying records, updating fields, invoking Flows, calling Apex, or triggering prompt templates. Each action is tied to a topic and represents a capability the agent can execute to accomplish a goal on behalf of the user.

Reasoning is the intelligent engine that ties everything together. The agent uses the Atlas Reasoning Engine to interpret a user's request, determine which topic applies, select the appropriate instructions, and decide which actions to execute in the correct sequence. This reasoning process allows the agent to think through problems, break them into steps, and respond dynamically rather than following rigid scripts.

Together, these components enable agents to understand context, make decisions, and take meaningful action. For administrators, mastering how topics, instructions, actions, and reasoning cooperate is essential for designing trustworthy agents that enhance productivity while keeping human oversight and security firmly in place.

When deciding between AI and rules-based automation in Salesforce, administrators must evaluate the nature of the problem being solved. Rules-based automation, such as Flows, Workflow Rules, Validation Rules, and Approval Processes, works best for tasks that are predictable, deterministic, and follow clear if-then logic. These scenarios have defined inputs and outputs, where the same condition always produces the same result. Examples include routing a case based on region, updating a field when a status changes, or sending an email when a record meets specific criteria. Rules-based automation is transparent, easy to audit, and reliable for compliance-heavy processes.

Artificial Intelligence, powered by Agentforce and Einstein, becomes appropriate when the problem involves ambiguity, natural language, pattern recognition, or predictions that cannot be captured through fixed rules. AI excels at tasks like summarizing long case histories, generating personalized responses, predicting lead conversion likelihood, classifying sentiment, or answering open-ended customer questions. These situations require interpreting unstructured data or making probabilistic judgments where outcomes vary based on context.

A practical approach is to combine both. Agentforce agents can handle conversational interactions and reasoning, while triggering deterministic actions through Flows for the parts requiring precision and control. For instance, an AI agent might interpret a customer request, then invoke a rules-based Flow to process a refund according to company policy limits.

Administrators should choose rules-based automation when consistency, auditability, and strict business logic matter most. They should choose AI when flexibility, language understanding, and adaptive decision-making add value. Cost, governance, and explainability also influence the choice, since AI outputs can be less predictable. By assessing whether a task needs a fixed answer or intelligent interpretation, administrators can select the right tool. Blending both strategies often delivers the strongest results, ensuring efficiency alongside trustworthy, controlled outcomes within the Salesforce platform.

The Einstein Trust Layer is a secure architecture built into Agentforce that protects sensitive data while enabling generative AI features across the Salesforce Platform. It includes several key components that administrators should understand.

Data Masking works by detecting and replacing sensitive information such as personally identifiable information (PII), payment details, and confidential business data before prompts are sent to a large language model (LLM). The system identifies these values and substitutes them with placeholder tokens. Once the LLM returns a response, the original values are restored (de-masked) so that users see accurate results while the third-party model never receives the actual sensitive content. This helps organizations maintain compliance and safeguard customer trust.

Zero Data Retention is a contractual and technical commitment ensuring that data sent to external LLM providers is not stored or retained after processing. When a prompt is submitted, the model generates a response and then the request and response are purged from the provider's systems. This means your company's prompts are not used to train external models, and no residual copies remain outside of Salesforce's secure environment. This gives administrators confidence that using generative AI does not expose data to long-term storage risks with vendors.

Toxicity Detection scans both the prompts submitted and the responses produced by the AI to identify harmful, offensive, biased, or inappropriate content. Each response receives a toxicity score, and this information is logged for auditing purposes. Administrators can review these scores to monitor the quality and safety of AI-generated outputs, helping ensure that content shared with customers and employees remains professional and appropriate.

Together, these features form a layered defense that allows businesses to adopt Agentforce and generative AI responsibly, balancing innovation with security, privacy, and ethical standards. Understanding them is essential for administrators managing trustworthy AI deployments.

In Salesforce Agentforce, agent security governs what data an autonomous or assistive agent can access and modify, ensuring it operates within the boundaries defined for its associated user context. Agents run under a specific user, often called the Agent User or a service account, and they inherit that user's permissions. This means an agent can only see and change data that the underlying user is permitted to access through the standard Salesforce security model.

Several layers control this access. First, profiles and permission sets determine object-level and field-level permissions, dictating which objects the agent can read, create, edit, or delete, and which specific fields are visible. Second, organization-wide defaults, role hierarchy, and sharing rules control record-level visibility, ensuring the agent respects who owns records and how they are shared across the org. Third, field-level security hides sensitive fields such as personal or financial details, so the agent cannot expose them even if it can view the record.

Agentforce also relies on Actions and Topics, which are curated capabilities that define what an agent is allowed to do. Administrators explicitly grant these actions, so an agent performs only the tasks that have been approved, such as looking up a case, updating a contact, or creating an opportunity. If an action is not assigned, the agent cannot perform it.

Data masking, encryption, and the Einstein Trust Layer add further protection, helping keep sensitive information secure during processing and preventing it from being retained inappropriately. Audit trails and monitoring allow administrators to review agent activity for compliance.

By configuring profiles, permission sets, sharing settings, field-level security, and assigned actions carefully, administrators ensure an agent sees and changes only appropriate data. This layered approach keeps Agentforce aligned with the same trust and governance principles that protect all Salesforce data.

In Salesforce Agentforce, Agent Users, Permission Sets, and Agent Access work together to enable and secure autonomous AI agents that operate within your org.

An Agent User is a special user account that represents the Agentforce agent inside Salesforce. When an agent performs actions such as querying records, updating cases, or executing flows, it does so under the identity of this Agent User. This approach ensures that all agent activity is tracked, auditable, and governed by the same security model that applies to human users. The Agent User has its own profile and can be assigned licenses and permissions like any standard user.

Permission Sets are used to grant the Agent User the specific capabilities it needs to function. Rather than modifying a base profile, administrators assign granular permissions through Permission Sets, which control object-level access, field-level security, Apex class execution, and access to specific features. By carefully scoping these Permission Sets, an administrator can follow the principle of least privilege, ensuring the agent can only perform tasks that align with its intended purpose. This modular approach makes it easier to adjust, audit, and revoke access as business needs evolve.

Agent Access refers to the overall configuration that determines what data, actions, and topics an agent is permitted to handle. This includes assigning the appropriate Permission Sets to the Agent User, defining which actions the agent may invoke, and setting boundaries around the records and knowledge sources it can reference. Proper Agent Access configuration protects sensitive data and helps maintain compliance.

Together, these three elements form the foundation of secure agent deployment. Administrators should regularly review the Agent User's assigned Permission Sets and monitor activity logs to confirm the agent behaves as expected, adjusting access whenever responsibilities change to keep the org protected and well governed.

Troubleshooting Agent permissions and access errors in Salesforce Agentforce involves systematically diagnosing why an AI agent cannot perform actions or retrieve data as expected. The most common cause relates to the permissions assigned to the agent's associated user or connected app. Every Agentforce agent operates under a specific user context, so administrators must verify that this user has the appropriate profile, permission sets, and permission set groups granting access to relevant objects, fields, and Apex classes.

Start by checking object-level and field-level security (FLS). If an agent fails to read or update records, confirm that the running user has Create, Read, Edit, and Delete permissions on the target objects, along with visibility to specific fields. Next, review sharing rules and record ownership, since role hierarchy and sharing settings determine which records the agent can access.

For agents invoking actions such as Flows or Apex, ensure the running user has execute access to those resources. A Flow may fail if the user lacks the 'Run Flows' permission or access to referenced objects. Similarly, Apex actions require access to the corresponding Apex class.

Agentforce also relies on topics and actions being properly configured and assigned. Verify that the agent has the correct topics enabled and that each action is activated and mapped to accessible resources. Use the Agent Builder testing panel to reproduce errors and examine debug logs, which reveal permission denials and specific failure points.

Additionally, check connected app settings, OAuth scopes, and Data Cloud permissions if the agent integrates external systems or grounding data. Reviewing setup audit trails helps identify recent changes that may have caused new errors.

A methodical approach checking user context, profiles, permission sets, sharing, FLS, and action access resolves most Agentforce permission issues efficiently, ensuring the agent operates securely within intended boundaries.

Enabling Einstein Generative AI and Agentforce in Setup is a foundational task for Salesforce Administrators who want to bring artificial intelligence capabilities into their org. To begin, an administrator must navigate to Setup and locate the Einstein Generative AI section, often found under the Einstein or AI settings. The first step involves turning on the Einstein Generative AI feature, which activates the platform's ability to use large language models for tasks such as content generation, summarization, and intelligent responses. Once enabled, the administrator should review and accept any applicable terms and conditions, as generative AI features may require acknowledgment of usage policies related to data handling and trust.

Next, to enable Agentforce, the administrator locates the Agentforce settings within Setup. Agentforce allows organizations to build autonomous agents that can perform tasks, answer questions, and assist users or customers. After toggling Agentforce on, the administrator can configure agent types, assign topics, and define the actions that agents are permitted to take. This ensures that agents operate along the boundaries set by the organization.

A key part of this process is verifying that the correct permissions and licenses are in place. Administrators should confirm that users who will interact with these features have the appropriate permission sets assigned. Additionally, reviewing the Einstein Trust Layer settings is recommended, since it governs how data is masked and secured when passed to AI models.

After enabling both features, testing is essential. Administrators can create a sample agent or generative prompt to confirm everything functions as expected. Monitoring usage through available dashboards helps track adoption and performance. By carefully following these steps, an administrator ensures that Einstein Generative AI and Agentforce are properly activated, secure, and ready to enhance productivity across the organization while maintaining compliance and trust standards.

Grounding agents in Knowledge Articles and CRM data is a foundational concept in Agentforce that ensures AI agents provide accurate, relevant, and trustworthy responses based on your organization's actual information. Grounding means connecting the AI agent to authoritative data sources so it generates responses rooted in real business context rather than relying solely on general training data.

Knowledge Articles serve as a curated library of verified content, including FAQs, troubleshooting guides, and policy documentation. When an agent is grounded in Knowledge, it retrieves and references these articles to answer customer or employee questions with approved, consistent information. Administrators must ensure articles are published, properly categorized, and assigned appropriate data categories so the agent can surface the correct content.

CRM Data grounding connects the agent to live records such as Accounts, Contacts, Cases, Opportunities, and custom objects. This allows the agent to personalize responses using real-time information, for example referencing a customer's open case status or recent order history. The agent respects the same security model, meaning field-level security, sharing rules, and permissions determine what data the agent can access on behalf of a user.

Retrieval Augmented Generation (RAG) is the underlying technique that combines these grounded sources with the large language model. When a query arrives, the system searches relevant Knowledge and CRM data, then supplies that context to the model to craft a grounded answer.

For administrators, key responsibilities include configuring data sources, maintaining data quality, setting up search indexes, and defining which objects and fields the agent can reference. Well-grounded agents reduce hallucinations, improve trust, and deliver responses aligned with company standards.

Ultimately, grounding transforms a generic assistant into a reliable, context-aware agent that reflects your unique business data, empowering better service and productivity while keeping security and accuracy at the forefront of every interaction across the platform.

Deploying Agents to Channels and Escalating to Human Reps are essential concepts within Agentforce that empower administrators to extend AI-driven assistance across customer touchpoints while maintaining quality service. When you deploy an Agentforce agent to channels, you are making the AI assistant available across multiple communication surfaces such as web chat, messaging apps (like WhatsApp and SMS), Slack, mobile apps, and embedded service experiences. Administrators configure these deployments through the Agentforce setup interface, connecting the agent to specific channels via the Messaging or Embedded Service Deployment settings. Each channel deployment can be customized with unique branding, greeting messages, and topic scopes, ensuring the agent behaves appropriately for its context. Proper deployment also involves assigning the correct actions, topics, and permissions so the agent can respond accurately to customer inquiries within that channel. Escalating to Human Reps is a critical capability that guarantees customers receive human support when the AI reaches its limits or when a situation requires empathy, complex judgment, or specialized authority. Administrators define escalation rules and criteria that trigger a handoff, such as low confidence responses, sentiment detection indicating frustration, explicit customer requests for a person, or specific topics flagged for human review. During escalation, the agent transfers the conversation along with full context, including chat history and customer details, so the human representative can continue seamlessly using tools like Omni-Channel routing and the Service Console. This preserves continuity and enhances customer satisfaction. Administrators should thoroughly test both deployment configurations and escalation paths, monitor performance through analytics and reporting, and refine topics and actions based on real interactions. By mastering these processes, a Platform Administrator ensures that Agentforce delivers scalable, efficient automated service while keeping humans available for cases that benefit from personal attention, creating a balanced hybrid support model that serves customers effectively across every channel.

Responsible AI in Salesforce and Agentforce centers on building trust through guardrails, human oversight, and clear limits on generative AI. Guardrails are protective mechanisms that keep AI behavior aligned with ethical standards, company policies, and legal requirements. In Agentforce, guardrails restrict what topics an agent can address, define allowed actions, and prevent harmful or inaccurate responses. Administrators configure these boundaries using topics, instructions, and scope definitions so agents stay focused on approved tasks. The Einstein Trust Layer supports guardrails by masking sensitive data, checking for toxicity, and grounding responses in trusted company data to reduce hallucinations.

Human oversight ensures people remain accountable for AI decisions. Rather than allowing full automation for critical outcomes, Salesforce encourages a human-in-the-loop approach where users review, approve, or edit AI-generated content before it reaches customers. This oversight is essential for high-stakes scenarios such as financial advice, legal matters, or sensitive customer communications. Administrators can design flows and permissions that require approval steps, giving employees final authority. This keeps organizations responsible for outcomes and allows corrections when AI produces unexpected results.

Generative AI limits acknowledge that these tools have boundaries. Large language models may generate plausible but incorrect information, reflect bias from training data, or struggle with context they were never given. Understanding these constraints helps administrators set realistic expectations and avoid overreliance. Grounding responses in verified Salesforce data, applying data privacy protections, and monitoring output quality all help manage these limitations.

Together, these principles form a framework for trustworthy AI adoption. Administrators must balance innovation with accountability, ensuring that automation enhances productivity while protecting customers and the business. By combining well-configured guardrails, meaningful human review, and awareness of generative AI limits, organizations can deploy Agentforce responsibly and maintain confidence in the technology they use every day for their teams.

Agent Builder is a core tool within Agentforce that allows Salesforce administrators to design, configure, and manage intelligent agents that assist users and customers. Two fundamental building blocks in Agent Builder are Topics and Actions, which define what an agent can understand and what it can accomplish. Topics represent the different areas of expertise or subject matter an agent can handle. Each topic groups related capabilities together, helping the agent recognize user intent and route conversations appropriately. For example, a service agent might have topics such as 'Order Management,' 'Returns,' or 'Account Support.' When configuring topics, administrators provide a clear name, description, and scope so the agent can determine which topic aligns with a user's request. Well-defined topics improve the agent's ability to respond accurately and stay focused on relevant tasks. Actions are the specific operations an agent performs to fulfill user requests within a topic. Actions can include retrieving records, updating data, creating cases, sending emails, or invoking custom flows and Apex. Each action has defined inputs and outputs, allowing the agent to gather necessary information and return meaningful results. Administrators assign actions to topics so that when a topic is triggered, the appropriate actions become available. Managing topics and actions involves creating, editing, testing, and refining these components to ensure the agent behaves as intended. Administrators should craft precise instructions and descriptions, since these guide how the agent interprets requests and selects the correct action. Testing is essential to confirm the agent responds correctly across various scenarios. Best practices include keeping topics distinct to avoid overlap, providing detailed action metadata, and iterating based on real user interactions. By thoughtfully managing topics and actions, administrators empower agents to deliver reliable, context-aware assistance, enhancing productivity and customer satisfaction while maintaining alignment with organizational goals and security requirements throughout the Salesforce environment.

In Agentforce, topics are the building blocks that define what an agent can handle. Writing and updating topic descriptions, scope, and instructions is essential for guiding how an autonomous agent behaves and responds to user requests. Each element serves a distinct purpose in shaping agent performance.

The topic description explains the overall purpose of a topic. It should clearly state what business area or task the topic covers, such as order management, case creation, or account inquiries. A well-crafted description helps the agent classify incoming user requests and route them to the correct topic, improving accuracy and relevance of responses.

The scope defines the boundaries of what the topic should and should not do. It clarifies the range of actions the agent may take within that topic. By setting a precise scope, administrators help the agent understand when a request belongs to a particular topic versus another, reducing confusion and keeping interactions focused on appropriate tasks.

Instructions provide detailed, step-by-step guidance on how the agent should behave when the topic is triggered. These are natural language directives that tell the agent how to respond, what tone to use, which actions to invoke, and how to handle specific scenarios or edge cases. Clear instructions ensure consistent, reliable, and business-aligned behavior.

When writing or updating these elements, administrators should use concise, unambiguous language and align content with business goals. Testing changes in the Agent Builder or Testing Center helps verify that the agent responds as expected. Regular updates keep topics relevant as business needs evolve.

As a Platform Administrator, mastering these components allows you to fine-tune agent capabilities, maintain governance, and deliver trustworthy automated experiences. Strong descriptions, well-defined scope, and precise instructions together form the foundation of effective, controllable, and high-performing Agentforce agents that meet organizational requirements.

Prompt Builder is a Salesforce tool that lets administrators create reusable AI prompt templates powered by generative AI within the Einstein/Agentforce platform. It has three core concepts: Prompt Template Types, Merge Fields, and Grounding.

Prompt Template Types define the purpose and context of a prompt. The main types include: Sales Email, which generates personalized emails tied to a record such as a Contact or Lead; Field Generation, which populates a specific field on a record with AI-generated content; Record Summary, which summarizes information about a record for quick review; and Flex templates, which are flexible and can be linked to multiple objects or used across various scenarios. Choosing the correct type ensures the prompt aligns with the business use case and available data.

Merge Fields allow you to insert dynamic Salesforce data into your prompt text. Instead of static wording, you reference fields from records, related objects, or flows using resource pickers. For example, a merge field might pull a customer's first name or an Opportunity amount so the generated output is personalized for each record. This makes templates reusable across many records while keeping content relevant.

Grounding is the process of supplying the large language model with trusted, relevant Salesforce data so responses are accurate and contextual rather than generic. Grounding sources include record fields, related lists, flows, Apex, Data Cloud, and retrieval augmented generation. By grounding a prompt in real CRM data, you reduce hallucinations and improve trust and quality of the output.

Together, these elements let admins build powerful, secure, and personalized AI experiences. You select a template type suited to the task, add merge fields for dynamic personalization, and apply grounding to anchor responses in reliable data. The Einstein Trust Layer further protects data privacy and masks sensitive information during processing, supporting responsible AI within Agentforce.

Installing and activating Prompt Templates and Agent Packages is a key skill for Salesforce Administrators working with Agentforce. Prompt Templates are reusable structures that define how generative AI produces content, such as summaries, emails, or field recommendations. Agent Packages bundle autonomous agents, their topics, actions, and supporting configurations for deployment across environments.

To install these components, administrators typically use managed or unmanaged packages distributed through the AppExchange or via installation URLs. Begin by navigating to Setup and locating the package installation link. During installation, choose the appropriate security level, granting access to admins only, all users, or specific profiles. Verify that all dependencies, such as required permission sets and Data Cloud connections, are satisfied before proceeding.

After installation, Prompt Templates appear within the Prompt Builder in Setup. Here, administrators can review the template type, such as Sales Email, Field Generation, or Flex, and inspect the grounding data that supplies context to the large language model. To activate a Prompt Template, open it in Prompt Builder, confirm the resource bindings and merge fields are correct, then click Activate. Testing using the preview panel helps confirm the output meets expectations before rollout.

For Agent Packages, once installed, the agents become visible in the Agentforce Studio or Agent Builder. Administrators must assign the necessary permission sets, connect the agent to relevant topics and actions, and configure the agent user. Activation involves reviewing the agent settings, enabling it, and publishing so end users can interact with it.

Post-activation, monitor performance through debug logs, testing centers, and analytics. Administrators should also manage version updates carefully, since upgrading packages may overwrite customizations. Following governance practices, such as documenting changes and validating in a sandbox first, ensures reliable deployment of both Prompt Templates and Agent Packages within the organization.

Light testing an Agent in Conversation Preview is a validation approach used within Agentforce to verify that your agent behaves as expected before deploying it to production or exposing it to real users. Conversation Preview is a built-in tool inside the Agent Builder that lets administrators interact with the agent in a simulated chat environment, sending sample messages and observing how the agent responds in real time.

During light testing, an administrator engages in quick, informal conversations with the agent to confirm that it triggers the correct topics, invokes the appropriate actions, and returns accurate, relevant responses. This helps you spot obvious issues early, such as an agent selecting the wrong topic, failing to call a needed action, or generating unexpected replies.

A key feature of Conversation Preview is the ability to inspect the reasoning behind each response. You can expand the conversation details to see which topic was matched, which actions were executed, what inputs and outputs were passed, and how the large language model interpreted the user's request. This transparency allows administrators to fine-tune instructions, topic classifications, and action configurations based on observed behavior.

Light testing is intended to be fast and iterative rather than exhaustive. It focuses on core happy-path scenarios and common user intents to confirm that essential functionality works. It is a complement to more rigorous, structured testing methods that cover edge cases, large volumes of test utterances, and comprehensive quality assurance.

By using Conversation Preview for light testing, Salesforce administrators can rapidly iterate on agent design, catch misconfigurations, and build confidence that the agent meets business requirements. This lowers risk when the agent moves toward broader deployment. Overall, light testing in Conversation Preview serves as an accessible, low-effort first checkpoint in the agent development lifecycle, enabling continuous refinement and improved end-user experiences across supported channels.

In Agentforce, managing the lifecycle of agents and prompts is a core administrative responsibility that ensures reliable and controlled deployment of AI-powered functionality. Activating an agent makes it live and available to interact with users, executing its assigned actions and topics based on its configuration. Before activation, administrators should thoroughly test the agent in a sandbox or preview environment to confirm it behaves as expected. Deactivating an agent takes it offline, which is useful when you need to pause its operation, perform maintenance, or make significant changes. A deactivated agent will not respond to requests, so administrators must communicate such changes to stakeholders to avoid disruption.

Versioning is a critical feature that allows administrators to maintain multiple iterations of an agent or prompt over time. When you modify an agent or a prompt template, Salesforce lets you create a new version rather than overwriting the existing one. This provides a safeguard, because you can revert to a previous version if a new change introduces unexpected results. Each version can be tracked, compared, and tested separately, giving teams confidence when rolling out updates.

For prompts specifically, versioning enables administrators to refine the instructions, grounding data, and output format iteratively. You can activate a particular prompt version to make it the one used in production while keeping earlier versions archived for reference. This supports a structured approach to continuous improvement.

Best practices include documenting the purpose of each version, testing new versions in a controlled setting, and coordinating activation timing with business needs. By carefully activating, deactivating, and versioning agents and prompts, administrators maintain governance, reduce risk, and ensure that AI capabilities align with organizational goals. This disciplined approach helps teams deliver consistent, trustworthy experiences while allowing room for experimentation and enhancement across the Salesforce platform.