Observe Phase Observability Tools

Agent Analytics

Agent Analytics is where you monitor your agent's performance over time. It aggregates data across all sessions and surfaces trends in a set of dashboards powered by Tableau Next. This is the starting point for any Observe phase question; look here first to understand what's trending before going to session-level tools.

Performance Insights

Performance Insights is the step between seeing a trend and diving into sessions. Before opening Agent Optimization, use it to break down any metric by subagent, intent, or action, so you know exactly where to look rather than reviewing sessions at random

See Agent Analytics for setup requirements and permissions.

Agent Optimization

When Analytics surfaces a pattern and you need to understand what's driving it, Agent Optimization is where you go next. It gives you a view into individual agent-user conversations — letting you filter, browse, and open specific sessions to see what actually happened. Two views:

• Processed Sessions — intent extracted, clustered, and quality scored. Available 4–5 hours after session close. Table columns: Session ID, timestamp, duration, outcome, custom scorer labels, intent summary, response summary, subagents triggered, actions triggered, intent tag, quality score.
• Unprocessed Sessions — active, recent, or still processing. Trace tab is available immediately — you don't need to wait for intent processing to start debugging.

Intent clustering is what separates pattern detection from session-by-session review. The Clustering Pipeline runs weekly and groups similar intents by semantic similarity once a cluster tag accumulates 10 or more matching intents. Once clusters form, you can filter by cluster to surface all sessions where users asked the same thing different ways — the fastest path to identifying systematic failure patterns rather than reviewing sessions one by one.

KPIs surfaced: Agent Impact and Agent Effectiveness outcome areas — session-level outcome, quality score, and intent patterns.

What it can't do: Aggregate metrics across sessions (use Analytics); sessions via email in real time aren't immediately reflected.

It surfaces session-level outcome, quality score, and intent patterns. See Agent Optimization for current session columns and filters.

Interaction Summary tab

When you open a session in Agent Optimization and select an intent, the Interaction Summary tab gives you a structured breakdown of what happened for that intent:

• Agent name, avg agent latency (ms), total interactions, intent duration
• Topics (subagents) triggered, actions triggered
• Intent tag, quality score (High/Medium/Low)
• Quality score reasoning — an LLM-generated explanation of why the quality score was assigned

Quality score reasoning is what converts a score into a direction. For a low quality score, the reasoning field often tells you directly what the evaluator found — wrong subagent, incomplete response, missed instruction, retrieval failure — in natural language. It's the fastest path from "quality is low" to "here's what to investigate," before you open the Trace tab.

It's not available until intent processing completes (4 to 5 hours post-session). For immediate debugging, start with the Trace tab.

Trace tab

Also accessible from a session in Agent Optimization, the Trace tab gives you a time-ordered record of every processing step — subagent identification, action selection, routing, variable updates, retrieval, and response generation. Each step shows a status indicator and timing, and nested steps are expandable for subagent delegation.

The Trace tab is the primary root cause analysis tool. Every other tool tells you what, like a metric moved, a quality score dropped, or a session failed. The Trace tab tells you where and why including which step failed, what the error message was, where time was spent, and what the retrieval returned.

Step types: intent handling, subagent routing, filters, retrievals, variable updates, and action execution. Each step has start and end timestamps and ErrorMessageText. For multi-agent (SOMA) deployments, expand any step where a subagent was invoked to drill into nested steps.

Working with errors: Query ErrorMessageText on AiAgentInteractionStepDmo__dlm via SOQL and filter out sentinel values (NOT_SET, None, empty string) to isolate real errors.

KPIs surfaced: Agent Health outcome area step-level latency, error location, and execution sequence.

Available immediately, even for unprocessed sessions. Don't wait for intent processing if you need to debug right now.

Agent Health Monitoring

Agent Health Monitoring detects silent agent failures like spiking error rates, high latency, or dropping response rates, and notifies you in near-real time before they compound into customer impact. Where Agent Analytics surfaces trends over time, Health Monitoring is designed for rapid response: set thresholds on the metrics that matter to your deployment, get notified when they're breached, and drill into session traces to find the cause.

These metrics are also visualized in the Health subtab of Agent Analytics.

Alerts

Each alert monitors one metric for one agent. The system evaluates alerts every minute against the previous 15 minutes of data and sends an email notification when a threshold is breached. After a notification fires, the same alert pauses for 30 minutes before it can notify again.

Notifications link the alert directly to the sessions that triggered it. Each notification shows the affected session so you can go straight to the trace without searching. See Monitor Agent Health for setup requirements and permissions.

Monitor Agent Guardrails

Monitor Agent Guardrails gives you a focused view of instruction adherence and task resolution — the Trust Layer outputs surfaced as dedicated monitoring metrics, separate from the broader Analytics view. For deployments where guardrail compliance is a tracked requirement — regulated industries, enterprise policy enforcement, brand safety — this gives you instruction adherence and task resolution trends in one place, separate from performance and quality signals.

KPIs surfaced: Agent Health outcome area — instruction adherence and task resolution.

See Monitor Agent Guardrails for current metrics.

Monitor Your Voice-Enabled Agent

For voice deployments, supervisors can monitor active calls in real time via the Agentforce tab in Omni Supervisor (Service Cloud console). This is distinct from the Voice tab in Analytics, which shows historical aggregate data — Omni Supervisor is for watching live activity, not reviewing past sessions.

The Agentforce tab in Omni Supervisor shows a list of agents on active calls and live call transcripts. It's accessed through the Service Cloud console, not Agentforce Studio.

KPIs surfaced: Agent Health outcome area — live call status and transcript, real-time not historical. See Monitor Your Voice-Enabled Agent for setup and supervisor configuration.

Data 360 Query Editor and Reports

For questions the product UI doesn't answer, the Data 360 Query Editor gives you direct SQL access to the underlying data. Query using DMOs — the semantic entry point to the data — across all four layers: STDM, Audit & Feedback, RAG Quality, and Agent Optimizer.

You can also build reports in Data 360 that pull relevant information from the applicable DMOs. To learn more, including required permissions, see Data 360 Reports and Dashboards.

Session Tracing Data Model (STDM) DMOs:

• AiAgentSessionDmo — session-level facts: channel type, session end type, related voice/messaging session IDs
• AiAgentSessionParticipantDmo — who joined: contact, lead, user, agent API name
• AiAgentInteractionDmo — one turn; includes TopicApiName (which subagent handled it)
• AiAgentInteractionStepDmo — one discrete step with start and end timestamps and ErrorMessageText
• AiAgentInteractionMessageDmo — actual input/output text

Audit & Feedback DMOs:

• GenAiGatewayRequest — the prompt and context sent to the LLM
• GenAiGatewayResponse — the LLM's response
• GenAiResponseGeneration — generation-level metadata
• GenAiFeedback — user feedback signals
• GenAiContentQuality — quality evaluation scores
• GenAiContentQualityCategory — content category classifications

RAG Quality:

• AIRetrieverRequest (DLO) — the retrieval query sent to the knowledge source
• AIRetrieverResponse (DLO) — the retrieved context returned
• AIRetrieverQualityMetric (DMO) — Answer Faithfulness, Answer Relevance, and Context Relevance scores

For Agent Optimizer objects, see Agent Optimization data model.

KPIs surfaced: All outcome areas — any KPI derivable from the raw data across all four infrastructure layers.

The next step is connecting them to surfaces. You've defined what success looks like for your agent, and you know what the toolchain can surface. Now, the work is building a plan that maps your outcome areas to the right signals, and the right signals to the right tools. That plan is what turns observability from a set of features into a repeatable practice..