Agent Analytics and Monitoring | 衡量、调试并持续改进 AI Agent 性能（练习操作步骤）

本文是《Agent Analytics and Monitoring | 衡量、调试并持续改进 AI Agent 性能》的练习操作步骤。阅读原文请见：Agent Analytics and Monitoring | 衡量、调试并持续改进 AI Agent 性能。

1. Launch Agent Analytics in Agentforce Studio

After you select the agent type, the Agent Analytics dashboard displays metrics for that agent type.
Use the Filter Bar to scope the data: select the agent, time window, channels, and modality (text/voice/both).
Select the Overview tab to see headline KPIs, period-over-period trend cards, and outcome charts.
Select a dimension tab (Effectiveness, Usage, Quality, Health, Trust, User Satisfaction, Voice) to view its KPI set and charts.
On the Effectiveness tab, review: Deflection Rate, Escalation Rate, Abandonment Rate, Engagement Rate, Success Rate, Task Resolution Rate.
Toggle between Metric Card view (large cards with graphics) and Table View (compact layout with all metrics at once).
Use the Date Filter to change chart granularity (Day/Week/Month).
Select the Usage tab to see: Unique Sessions, Unique Interactions, Unique Users, Average Interactions Per Session.
Select the Quality Scores tab to see dimension scores (Quality Score 1–5, Answer Faithfulness 0–1, Answer Relevance 0–1, Context Relevance 0–1).
Select the Health tab to see: Error Rate, Session Duration (seconds), Agent Interaction Latency (seconds).
Select the Trust tab to see: Adherence Response Rate, Average Agent Toxicity Score (0–1).
Select the Voice tab to see: Interruption Rate.

Select the Performance Insights tab.
In the Breakdowns panel, choose a segment type: Subagents, Intents, or Actions.
Use Select Subagent (or the equivalent selector) to filter to a specific segment, or leave it set to All.
In Select Metric, choose the measure you want to compare.
Review the horizontal bar chart. Each row represents a segment, with bar length representing the metric value and color coding indicating performance level (green=favorable, orange/red=needs attention).

Log in to Agentforce Studio.
Navigate to Observe & Optimize | Analytics.
Select the Service Agent type and change the Timeframe in the Filter Bar to the Last 30 Days.
On the Agent Performance tab, review the Effectiveness dimension tab.
Check: Deflection Rate, Escalation Rate, Abandonment Rate, Engagement Rate, Success Rate.
Study the Aggregated Effectiveness Metrics trend chart over the 30-day period.
Check the Session Outcome stacked bar chart at the bottom (Deflected, Escalated, Abandoned, Ambiguous).
Switch to the Performance Insights tab.
In Breakdowns, select Segment type: Subagents, Metric: Deflection Rate.
Review the horizontal bar chart — each subagent shows a colored bar (green/orange/red).
Change the metric to Abandonment Rate and compare results.
Identify the underperforming subagent (highest abandonment rate, lowest deflection rate).
Write down action items: investigate subagent instructions, check Quality scores, check Health metrics (latency), meet with stakeholders.

Use the Filter Bar to scope the data: agent, timeframe, channels, modality, session outcome, quality score, subagent, intent tag.
Select the Processed Sessions tab to see sessions that have completed intent extraction and clustering.
- Intent Pipeline: Analyzes closed sessions and extracts intents (typically takes 4–5 hours after session closes).
- Clustering Pipeline: Groups similar intents across sessions, assigns cluster tags (runs weekly).
Select the Unprocessed Sessions tab to see sessions still active, recently closed, or not yet processed.
Review the Sessions table columns: Session ID / Timestamp, Session Duration, Session Outcome, Custom Scorers, Intent Summary, Response Summary, Subagents, Actions, Intent Tag, Quality Score.

Click the Session ID of the conversation you want to investigate.
The left panel shows the full conversation log with:
- Timestamps on every message
- Color-coded bubbles (user=right, agent=left)
- Quality badge on agent responses (High/Low)
- Agent completion time for each response
The right panel has two tabs: Interaction Summary and Trace.
In Interaction Summary, review:
- Agent Name, Average Agent Latency, Total Interactions, Intent Duration
- Topics Triggered (subagents invoked, color-coded)
- Intent Tag (cluster tag), Actions Triggered
- Quality Score (color-coded High/Medium/Low)
- Quality Score Reasoning (LLM explanation — most actionable field for knowing what to fix)
In the Trace tab, review the time-ordered record of every processing step:
- Topic (subagent) identification
- Actions (labeled with action type and icon)
- Agent routing steps, Variable updates, Retrieval steps, Response generation
- Each step shows a status indicator (green check=success) and timing
- Expand nested trace steps to drill into subagent delegation

From Agent Analytics, note the underperforming subagent (e.g., Cancellation & Rescheduling).
Check the Overview | Quality dimension to see overall and per-subagent quality scores.
Check Performance Insights | Subagents with Escalation Rate metric.
Review quality sub-scores: Answer Faithfulness, Answer Relevance, Context Relevance.
Navigate to Sessions & Intents.
Filter by:
- Subagent: (the underperforming one)
- Session Outcome: Escalated
- Quality Score: Low (1.0–3.0)
Open the top 5–10 escalated sessions.
For each session:
- Read the Chat Session Log to understand what the user asked and how the agent responded.
- Check the Interaction Summary tab for Quality Score Reasoning.
- Check the Trace tab to see which knowledge articles were retrieved and what actions were executed.
Identify patterns across sessions:
- Are there knowledge gaps? (multiple overlapping articles, missing exception scenarios, ambiguous policies)
- Are the subagent instructions clear enough?
Implement fixes:
- Consolidate fragmented knowledge articles into a single comprehensive article.
- Archive old articles so the agent retrieves only the new one.
- Update subagent instructions to handle edge cases.
Test the new version in Agentforce Testing Center.
Deploy and monitor for a few weeks.
Verify improvements in metrics (Quality Score, Answer Faithfulness, Escalation Rate).