Agentforce Considerations
Carefully review the support, limits, and considerations for AI agents.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Performance, Unlimited, and Developer Editions. Required add-on licenses vary by agent type. |
Important Starting June 17, 2025, Agentforce
(Default) will not include new features or improvements and isn’t available in new
Salesforce environments. We recommend migrating to Agentforce Employee agent for continued
enhancements and support. If you plan to transition to Agentforce Employee agent or make
related license changes, complete the migration of your existing Agentforce (Default) agents
in advance to avoid potential disruption in agent availability. If you can't complete the
migration in advance, you can create new agents after the transition, which can involve some
downtime. See Migrate from Agentforce (Default) to Agentforce Employee
Agent.
Agents Support
- Agentforce is included as a Covered Service in the Einstein Platform and Agentforce SOC 2 and SOC 3 reports. Agentforce is HIPAA eligible and covered under the Salesforce Business Associate Addendum Restrictions, and has also achieved ISO 27001, 27017, and 27018 certifications.
- The Agentforce panel is available in Lightning Experience, the Salesforce mobile app, the Field Service mobile app, and Sales Cloud Everywhere.
- The Agentforce platform supports OpenAI GPT-4o for reasoning engine calls. To use Anthropic on Amazon Bedrock as the model provider for Agentforce instead, see Select Agentforce Model Option. Agent actions can make calls to other predefined LLMs. Bringing your own model isn’t supported, but custom actions that execute prompt templates can use any Salesforce-managed model. See Large Language Model Support.
- Agentforce supports custom and standard User Interface API-supported objects. Object support for individual standard agent actions can vary. See the reference documentation for an action for more information.
- Currently, only Agentforce Service Agent can connect to enhanced Messaging channels and Bring Your Own Channel. See Considerations for Agentforce Service Agent.
- Agents have limited API support. See Agent API Developer Guide.
- Agents aren't available for Government Cloud.
Agents Languages
- Language support differs by agent type. For more details, see the considerations for your agent type.
- Language settings apply to LLM-generated messages only.
- Subagent and action instructions, including input and output instructions, are supported in English only. If you provide them in a different language, the agent's reasoning engine may not perform as expected.
- Language and tone settings don’t affect the output of agent actions that have a specified language and tone, such as Draft or Revise Email. Some action inputs are available in English only.
- System messages, such as the agent's welcome message, are translated into the end user's language only when the end user's language is selected as an additional language in the agent's language settings. To test system message translation, change the End User Language preview condition in the conversation preview panel.
- If the inputs or outputs of agent actions or associated prompt templates are in an unsupported language, those actions may not function properly.
- Language support for toxicity differs from the supported languages for Agentforce. See Einstein Trust Layer Region and Language Support.
- Some LLMs generate responses that are close variants but not exact matches for the requested locale. For example, if a user requests the French (Belgium) locale, which isn’t supported, some LLMs generate a response in the French (France) locale.
Agents Limits
- For best performance, we recommend assigning no more than 10 actions to a subagent, and 10 subagents to an agent.
- You can have up to 100 agents and, within that 100, you can have one version of the default Agentforce agent. Einstein bots count towards the limit of agents.
- If your licenses limit the number of a specific agent type that you can create and you reach that limit, the agent type is disabled in the guided setup.
- Agent actions time out after 60 seconds.
- Reasoning engine requests time out after 30 seconds.
- Agent action outputs over 65,000 characters are truncated.
Considerations for Agent Conversations
- Agents stream LLM responses in conversations as the responses are received. Certain Einstein Trust Layer features are applied only to the final agent response. If a hallucination is detected after the response has started streaming, the response is deleted and a new response is generated.
- Agents are optimized for questions or requests that are related to a specific topic, rather than open-ended questions that are open to interpretation.
- Deactivating an agent interrupts any ongoing user conversations. Users aren’t notified that an agent is deactivated. If the agent has system messages, it sends the system error message as a response to any user messages. If the agent is deployed to the Agentforce panel, the panel is available until the user closes it, and it can’t be reopened until the agent is active.
- In general, agents support one intent per utterance, so users can’t make requests that span multiple intents. Users can break more complex requests into individual messages. However, the Draft or Revise Email action can combine actions to address multiple user intents.
- Agent conversations aren’t personalized based on past sessions. See Agents Limits.
- Agents support small talk, but topics and number of turns are limited to ensure trust and prioritize enterprise use cases. Great for CRM! Awkward at parties.
Considerations for Custom Assets
- When you create a custom action, you can’t specify how the output appears in an agent conversation. The output is formatted automatically.
- Custom actions can’t accept a list and object as input. A fix is on the way.
- Custom actions that reference a prompt template can use any Salesforce-managed model. See Large Language Model Support.
- Custom actions that reference a custom prompt template support only these inputs and outputs:
- Input: record information object type inputs (for example, the opportunity to
summarize)
Note Object input types are currently only partially supported due to inconsistent behavior. A fix is on the way. In the meantime, we recommend using a flow to wrap a prompt template and creating a custom action from the flow. This scenario requires detailed action instructions, so test and iterate on your instructions carefully. - Outputs: text type outputs (for example, a two-sentence summary of the opportunity)
- Input: record information object type inputs (for example, the opportunity to
summarize)
- Testing or previewing custom Lightning types for Agentforce Service agent in Agentforce Builder is not supported. To test and preview custom Lightning types for Agentforce Service agent, navigate to the Embedded Service Deployments Settings page and click Test Enhanced Web Chat.
- Agent custom actions support only custom and standard User Interface API-supported objects.
- Custom actions that reference an Apex class or flow support only primitive data types. For example, if a flow includes a picklist, currency, or collection resource, you can't use that flow as a reference action for a custom action.
- Loading text for agent actions can be tested in preview conversations for active agents only.
- When you remove a custom subagent or action from an agent, it's deleted and can't be restored.
See Also
Did this article solve your issue?
Let us know so we can improve!
