Agentforce Observability Infrastructure

Agentforce Session Tracing

Agentforce Session Tracing tracks the agent activity layer. It captures every agent turn, reasoning step, action, message, and error for every session, and stores them in structured objects in Data 360. See Agentforce Session Tracing.

The Session Tracing Data Model (STDM) has five objects organized in a nesting hierarchy:

• AiAgentSession — the top-level record for each session. The field SessionEndType, which maps to deflected, escalated, and abandoned sessions.
• AiAgentSessionParticipant — who was in the session: contact, lead, user, or agent API name.
• AiAgentInteraction — one turn within the session. The field TopicApiName identifies which subagent handled the turn.
• AiAgentInteractionStep — one discrete step within a turn: tool call, Apex call, Flow call, or LLM call.
• AiAgentInteractionMessage — the actual input/output text for each turn.

These objects feed the Effectiveness, Usage, and Health tabs in Agent Analytics, and the session list and Trace in Agent Optimization. All five objects are directly queryable via Data 360 Query Editor for custom aggregations and joins. See Data Model for Agentforce Session Tracing for the full object reference.

Agent Platform Tracing

Agent Platform Tracing extends data collection to the service layer, what happened inside the Agentforce infrastructure to execute the decision Session Tracing records. Where Session Tracing tells you what the agent decided and said, Agent Platform Tracing tells you which service ran, for how long, and with what status. It captures service-level telemetry across Atlas reasoning engine, Apex, Flows, Prompt Builder, Invocable Actions, AI Gateway, LLM Gateway, and DC Query Federator. Data is stored in a single DMO, separate from STDM, and joinable via TelemetryTraceId.

• ssot__TelemetryTraceSpan__dlm — parent/child span relationships, duration in nanoseconds, operation names, service names, and status codes.

Together, this gives you a distributed trace across the full Agentforce ecosystem. See Agent Platform Tracing.

Audit and Feedback

Audit and Feedback tracks the LLM Gateway layer. It captures the inputs and outputs of every LLM call, including prompt text, response text, trust evaluations, guardrail outputs, and user feedback signals. While Session Tracing tells you what the agent did, Audit and Feedback tells you whether the LLM response was accurate, safe, and grounded in your knowledge source.

• GenAIGatewayRequest — the prompt and context sent to the LLM
• GenAIGatewayResponse — joins the gateway request to its generation output
• GenAIGeneration — the LLM-generated response text, including masked and demasked versions, and generation parameters
• GenAIFeedback — user feedback signals like thumbs up/down.
• GenAIContentQuality — whether an input/output is safe/unsafe, including toxicity detection
• GenAIContentCategory — Trust Layer detector results by type: InstructionAdherence, TaskResolution, TOXICITY, PROMPT_DEFENSE

Audit and Feedback data surfaces in Quality Scores, Trust, and User Satisfaction in Agent Analytics, and the quality score reasoning field in Agent Optimization. See Agent Analytics and Set Up Einstein Generative AI Audit and Feedback for current metrics.

Knowledge/RAG Quality Data and Metrics

Knowledge/RAG Quality Data and Metrics adds retrieval quality scoring to the Audit and Feedback pipeline. It captures how relevant the query was to the task, how faithful the response was to the retrieved knowledge source, and how relevant the retrieved knowledge source was to the query.

Key objects:

• GenAIRetrieverRequest — the retrieval query sent to the knowledge source
• GenAIRetrieverResponse — the retrieved context returned
• GenAIRetrieverQualityMetric — Answer Faithfulness, Answer Relevance, and Context Relevance scores

It populates the retrieval quality metrics in Quality Scores in Agent Analytics, which signals whether the knowledge source is returning the right content, separate from the LLM's response quality. See Set Up Knowledge/RAG Quality Data and Metrics for setup steps.

Trust Layer

The Trust Layer is the compliance scoring layer. It generates instruction adherence and toxicity scores from LLM Gateway outputs. Trust Layer scores are stored in GenAIContentCategory and GenAIContentQuality, the same objects from Audit and Feedback.

It surfaces in the Trust in Agent Analytics, giving you visibility into whether the agent is following its instructions and staying within acceptable content bounds. See Einstein Trust Layer for setup steps.

Each layer powers a different part of the observability toolchain. The more layers you enable, the more complete your view of agent behavior and the more of the Agent Development Lifecycle you can navigate with evidence.