[object Object]

What It Is

Agentforce 3 Command Center tracks every production agent — performance, error rate, escalation rate, adoption, cost. Real-time alerts for regressions. AI-driven recommendations for tuning topics and actions. Generally available August 2026, included with every Agentforce SKU at no additional license cost (you pay for Data Cloud storage of the traces). The Command Center is the operational layer; without it, scaling beyond a single Topic is guessing.

Session Tracing

Every agent interaction traces via OpenTelemetry. Spans capture topic classification, tool calls, grounding fetches, response generation, Trust Layer redactions, and downstream Apex callouts. Traces feed into Data Cloud for long-term analysis and can be exported to any OTLP-compatible backend (Datadog, Honeycomb, Grafana Tempo) via the standard collector. You can finally answer “why did the agent do that” without guessing.

Span hierarchy for one turn:
  agent.session
    \_ topic.classify       (model: atlas-router, latency_ms, confidence)
       \_ retrieval.fetch    (k=8, vector_index, chunks_returned)
       \_ action.invoke      (action_id, input_hash, output_hash)
          \_ apex.callout    (callout_endpoint, status, duration)
       \_ response.generate  (model, prompt_tokens, completion_tokens, cost_usd)

Each span carries a trace_id that ties the entire conversation together — useful when a user reports “it was wrong yesterday at 3pm” and you need to reconstruct exactly what the agent saw.

Operational Metrics

Watch topic misclassification rate, action failure rate, handoff rate, cost per conversation, and p95 latency. Publish these weekly. Threshold alerts catch drift — a spike in escalations often precedes a user complaint by 24–72 hours.

Recommended starting alerts:
  topic_misclass_rate     > 8%   over 1h window
  action_fail_rate         > 3%   over 30m window
  handoff_rate              up by 25% week-over-week
  cost_per_conv_p95        > 2x baseline over 1d
  p95_latency_ms           > 6000 over 15m window

Setup Priorities

Enable tracing before scaling agent traffic. Connect the Data Cloud destination early so historical data accumulates from day one. Define alerts with realistic thresholds — too sensitive and alert fatigue kills signal; too loose and you discover problems through Salesforce CSAT surveys instead of dashboards. Wire the alerts into the Slack channel where the agent owner actually reads, not a generic ops channel.

Common Failure Modes

  • Sampling traces to “save cost” — the bottom 10% of conversations are exactly the ones you need full fidelity on. Sample uniformly or not at all.
  • Treating handoffs as failures. A high handoff rate on a clearly-scoped low-confidence path is a feature, not a bug.
  • Ignoring cost p99 because the average looks fine — runaway loops show up in the tail first.

What Changed in 2026

The August 2026 GA added native session replay (visual playback of the agent UI alongside the trace), built-in regression diff (compare metrics across agent versions), and a Recommendations panel that uses Atlas to suggest Topic merges, Action consolidations, and Instruction edits based on observed failure patterns.

What to Do This Week

Enable Command Center tracing on your highest-volume production agent and configure five baseline alerts on misclassification, action failure, handoff rate, cost, and latency.

[object Object]
Share