Agentforce¶
TODO(David)
Write this page. Stub created for structure only.
TODO
Agents, topics and actions, prompt templates, Einstein and GenAI settings; review standards for agent behaviour changes
Source (Well-Architected, Compliant / ethical AI; see standards-sources memory)
Representative datasets and grounding-document quality (stale or contradictory RAG sources produce hallucinations); human-at-the-helm escalation for high-risk or low-confidence cases; bias, explainability, and robustness assessments plus ethical red teaming; model cards and AI-regulation readiness (EU AI Act); monitor for drift and fairness; Einstein Trust Layer
Standards for Agentforce and generative AI metadata: agents, topics, actions, and prompt templates, and how changes to agent behaviour are reviewed.
Observability¶
Research note (2026-09-28)
Agent Health Monitoring was announced as GA for Spring 2026 (Salesforce news, 20 Nov 2025) but described as closed beta in a Salesforce blog (18 Dec 2025); confirm current availability before stating it.
Agents are observed with the same seriousness as code. Agentforce Observability is Salesforce's observability layer for agents (announcement):
- Agentforce Session Tracing records every agent interaction in the Session Tracing Data Model, stored in Data 360: inputs, responses, reasoning steps, LLM calls, and guardrail checks.
- Agent Analytics tracks usage and effectiveness across all deployed agents, including escalation, deflection, and abandoned sessions.
- Agent Optimization gives step-by-step visibility into individual interactions, clusters sessions by intent, and scores their quality.
- Agent Health Monitoring watches error rate, latency, and escalation rate in near real time, with custom thresholds and alerts.
The standard:
- Session tracing is enabled for every agent in production, so any agent response can be traced back to the inputs, reasoning, and actions that produced it.
- A change to an agent's behaviour (its topics, actions, instructions, or prompt templates) is reviewed against its analytics before and after deployment, in the same way a code change is reviewed against its tests.
- Agent health alerts reach the same people, through the same channels, as every other production alert; see Observability.
- Session traces contain customer conversations, so they are classified and retained as customer data, not as diagnostic logs.