Managing Agentforce Offline Experimentation

When to Use This Skill

Execute this workflow when a user wants to test how a different LLM model would handle existing Agentforce agent conversations. Fetches session traces from Data Cloud STDM, replays prompts via the SFAP AI Gateway REST API with structured tool calling, scores across four metrics (topic agreement, function name accuracy, function argument accuracy via LLM-as-judge, text response quality via LLM-as-judge), and produces CSV + HTML reports with SAFE/REVIEW/UNSAFE verdict badges. Requires Data Cloud with Agentforce Activity data streaming enabled, an authenticated org, and a target model ID (llmgateway__ prefix for SFAP replay). Estimated runtime: 10-30 minutes for 10 sessions.

Workflow

discover-data-space → verify-stdm-availability → discover-agents → resolve-agent-label → list-available-models (if user needs help choosing) → find-sessions → get-conversation-details → get-llm-step-details → obtain-sfap-token → replay-turn-prompts → [score-exact-matches, judge-function-args, judge-text-responses] → generate-csv-report → generate-html-report. WORKFLOW GATE: Do not proceed past Phase 0 until org alias (auto-discover from authenticated orgs), agent API name (auto-discover from STDM), and target model ID (ask user) are all confirmed. SFAP JWT expires after ~2 hours. For 20+ session replays, refresh the token between batches. Each SFAP replay call takes ~1-3 seconds; keep under 20 sessions for reasonable wall-clock time (~10-30 minutes). The scoring and judging steps (Phase 3) can run in parallel once replay completes. Exact-match scoring is instant; LLM judge calls add ~1-2 seconds per turn.

Critical Constraints

Preconditions: