To ensure a custom action in Agentforce performs as expected, the AI Specialist must focus on Action Instructions. Here's why:
* Action Instructions define the logic, parameters, and steps the AI should follow to execute the action.
They include:
* How input data is processed.
* API calls or Apex invocations.
* Conditional logic (e.g., decision trees).Testing and iterating on these instructions ensures alignment with the intended workflow. For example, incorrect API endpoint references or misconfigured parameters in the instructions will cause failures.
* Action Input (Option A) refers to the data provided to the action. While validating input formats is important, inputs are static once defined. The primary issue lies in whether the instructions correctly use the inputs.
* Action Name (Option B) is a descriptive label and does not affect functionality.
Salesforce Documentation Support:
* Salesforce Einstein Bots & Custom Actions Guide highlights that Action Instructions are where the
"core logic" resides, requiring rigorous testing (Source: Einstein Bots Developer Guide).
* Trailhead Module "Build Custom Actions for Einstein Bots" emphasizes refining instructions to handle edge cases and validate outputs (Source: Trailhead).
By iterating on Action Instructions, the AI Specialist ensures the action's logic, integrations, and error handling are robust.

When a custom Agent action is not being prioritized correctly by the planner service, the root cause is often missing or improperly defined action dependencies. The planner service determines the execution order of actions based on dependencies defined in the action instructions. To resolve this, the AI Specialist must explicitly specify dependent actions using their API names in the custom action's configuration. This ensures the planner understands the sequence in which actions must be executed to meet business logic requirements.
Salesforce documentation highlights that dependencies are critical for orchestrating workflows in Einstein Bots and Agentforce. For example, if Action B requires data from Action A, Action A's API name must be listed as a dependency in Action B's instructions. The Einstein Bot Developer Guide states that failing to define dependencies can lead to race conditions or incorrect execution order.
In contrast:
* Profiles or custom permissions (B) control access to the action but do not influence execution order.
* LLM model provider and version (C) determine the AI model used for processing but are unrelated to the planner's sequencing logic.

To addresssecurity complianceconcerns and provide visibility into theprompt text sent to the LLM, how it ismasked, and themasked response, the AI Specialist should recommend enabling theaudit trail in the Einstein Trust Layer. This feature captures and logs the prompts sent to the large language model (LLM) along with the masking of sensitive information and the AI's response. This audit trail ensures full transparency and compliance with security requirements.
* Option A (Einstein Shield Event logs)is focused on system events rather than specific AI prompt data.
* Option B (debug logs)would not provide the necessary insight into AI prompt masking or responses.
For further details, refer toSalesforce's Einstein Trust Layer documentationabout auditing and security measures.