Section 6 · Governance and Observability · 10% of exam
6.3 Identifying Opportunities to Optimize Agent Behavior
Exam objective
Explain how administrators use session patterns, intents, traces, and quality signals to identify meaningful opportunities to improve an agent.
Look for repeated friction
Optimization begins with a pattern, not an assumption. Review unresolved sessions, low-quality interactions, frequent escalations, repeated clarifying questions, failed actions, or user feedback. A single unusual conversation may not justify a broad design change, but a cluster of similar sessions can reveal a recurring problem. Agent Optimization can organize sessions by intent and quality, then help surface repeated gaps. Compare the user’s request with what the agent understood, which subagent or topic it chose, which actions ran, what data was retrieved, and how the session ended. This creates a concrete explanation for the observed outcome. Find the layer causing the problem
A poor answer can come from multiple layers. If the wrong topic is selected, inspect classification descriptions, scope, and routing. If the correct topic is selected but an action is missing or wrong, inspect action instructions, filters, variables, inputs, and permissions. If the action succeeds but the response is inaccurate, review its outputs, grounding source, retrieval results, and final instructions. If an answer is correct but slow, use the trace’s step timing to locate expensive or delayed work. For voice, assess whether poor audio or speech recognition may have changed the input. Fix the underlying cause instead of rewriting unrelated instructions.
Use intent and quality evidence
Agent Optimization’s Sessions & Intents view can show intent summaries, response summaries, intent tags, quality scores, and actions. Processed sessions can be grouped to expose recurring request patterns; unprocessed sessions may be available before analysis completes. Drill into a specific session and use Interaction Summary for an overview, then Trace for the time-ordered processing steps and errors. When working with an intent cluster, review representative conversations rather than assuming every session in a group failed for the same reason. Confirm that the source data and evaluation signal support the diagnosis. An optimization suggestion is a hypothesis to examine, not an instruction to apply without review.
Prioritize practical fixes
Prioritize issues by frequency, user impact, risk, and how broadly a change may affect other behaviors. A missing source for a common policy question may merit a knowledge update; a sensitive action exposed before verification may need an access-control fix even if it is rare. Prefer a targeted change that addresses the evidence. If the remedy affects routing, actions, prompts, data, or permissions, add the failing pattern to the test suite and check nearby scenarios for regressions. Compare the results before and after the change. Keep an issue log with the observed behavior, evidence, root cause, fix, and test outcome so the team can learn from recurring gaps.
Diagnose in order: user intent, routing, instructions, variables and filters, action execution, retrieved data, response quality, and latency.
Agent Optimization can surface patterns or suggestions, but an administrator must validate the evidence and decide whether a change is appropriate.
Extra concept: Prioritize a low-frequency issue when its consequence is serious; volume alone is not a sufficient measure of risk.
Official references:
- Agent Optimization https://help.salesforce.com/s/articleView?id=ai.generative_ai_optimize.htm&language=en_US&type=5
- Analyze Data with Agent Optimization https://help.salesforce.com/s/articleView?id=ai.generative_ai_optimize_use.htm&language=en_US&type=5
- Understand Agentforce Observability KPIs and Tools https://help.salesforce.com/s/articleView?id=ai.agentforce_understand_observability_kpis_and_tools.htm&language=en_US&type =5
Key takeaways
- Use repeated unresolved sessions, quality signals, feedback, and intent patterns to identify opportunities.
- Trace the interaction to the layer that caused the issue before changing the agent.
- Prioritize by impact and risk, then add the issue to regression tests and measure the result.
