Agentforce Coworker Data in Data 360

Agentforce Coworker stores data in Data 360 as data model objects (DMOs) in two groups: LLM request data and agent session data. Learn which DMOs are populated for Agentforce Coworker and how Agentforce Coworker prompts are structured.

REQUIRED EDITIONS
Available in:	Enterprise, Unlimited, and Agentforce 1 Editions

LLM Request Data: Einstein Audit and Feedback DMOs

The Einstein Audit and Feedback DMOs show what the large language model (LLM) did. They store each request that Agentforce Coworker sends to the LLM, the response, token counts, Einstein Trust Layer masking and toxicity results, and user feedback. Each record is scoped to a single LLM request. In these DMOs, Agentforce Coworker records have the feature value Agentforce Coworker.

All Einstein Audit and Feedback DMOs are populated for Agentforce Coworker.

DMO	Contains
GenAIGatewayRequest	The LLM request, including the prompt, model, provider, tokens, feature, and timestamp
GenAIGatewayResponse	The LLM response
GenAIGeneration	The generated response text. When data masking is on, the masked response text isn’t stored.
GenAIGatewayRequestTag	Key-value tags on the request
GenAIGtwyRequestMetadata	Tool-call metadata
GenAIFeedback	Thumbs-up and thumbs-down feedback. Copy events aren’t recorded as feedback.
GenAIFeedbackDetail	Free-text feedback

For details on these DMOs and how to turn on data collection, see Einstein Audit and Feedback Data.

Agent Session Data: Agentforce Session Tracing DMOs

The Agentforce Session Tracing DMOs show how the agent ran. They group sessions, participants, turn-by-turn interactions, reasoning steps, and messages under one session ID.

Not all Agentforce Session Tracing DMOs are fully populated for Agentforce Coworker.

DMO	Contains	Status for Agentforce Coworker
AiAgentSession	The session container and IDs for one chat, from start to close	Populated
AiAgentInteractionMessage	User and agent message text	Populated
AiAgentGenerativeAiUsage	Token and generation usage	Populated
AiAgentSessionParticipant	The session participants	Partially populated. User IDs are populated. Agent IDs aren’t. Use “ssot__AiAgentApiName__c” = ‘SearchAgent’
AiAgentInteraction	The turns in a session	Not populated
AiAgentInteractionStep	Reasoning steps, tool calls, durations, and action logs	Not populated

Because interaction steps aren’t populated, you can’t use session tracing data to analyze Agentforce Coworker execution flow, tool-call success, or step timing. To review what Agentforce Coworker sent to the LLM, use the Einstein Audit and Feedback DMOs instead.

For field-level details, see Generative AI Session Trace Data Model and About Agentforce Session Tracing.

Types of LLM Calls That Agentforce Coworker Makes

Agentforce Coworker makes several types of LLM calls. All of them are recorded with the feature value Agentforce Coworker in the GenAIGatewayRequest DMO, but each type has a different prompt structure.

Primary conversational calls: The exchange between the user and Agentforce Coworker. Most audit trail queries target this call type.
Delegated agent calls: When Agentforce Coworker hands a task to an Agentforce employee agent, the employee agent’s LLM calls are also recorded. These prompts start with role: assistant and the employee agent’s persona.
Internal helper calls: Sub-calls that Agentforce Coworker makes to specialized workers, such as a search and retrieval helper.
Utility calls: Internal calls for safety guardrails, entity classification, and confirmation captions. These calls have minimal system and user prompts.

Queries that extract parts of the prompt return reliable results only for primary conversational calls. For other call types, they can return unexpected fragments or empty results.

Structure of a Primary Conversational Prompt

Agentforce Coworker sends the LLM the full compiled prompt, not only the user’s question. In the GenAIGatewayRequest DMO, the prompt field for a primary conversational call contains these sections, in order.

System instructions: The Agentforce Coworker persona, tone, and behavioral rules. This section is the same for every user and every request.
Static context: The user profile and org information.
## Dynamic Context: User context added at runtime, such as identity, recently viewed records, the active workspace, and session state. This section makes each prompt unique to the user and the moment.
### Agent Catalog: Where Agentforce Coworker uses the dynamic context to decide which agent or tool to call to retrieve data.
role: user: The generalized, rewritten version of the user’s question. This text is the closest record of what the user asked.

Use the ## Dynamic Context, ### Agent Catalog, and role: user markers to extract specific sections of the prompt. For example queries, see Review Agentforce Coworker Prompts and Responses in Data 360.