agentforce-offline-experimentation-run

Imported from forcedotcom/sf-skills (skills/agentforce-offline-experimentation-run/SKILL.md). Install upstream with npx skills add forcedotcom/sf-skills --skill agentforce-offline-experimentation-run. Copyright stays with the author.

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 → 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:

• Data Cloud is provisioned and enabled in the org (check: DataCloudProvisioned)
• Einstein LLM / GenAI Gateway is configured in the org (check: EinsteinLlmEnabled)
• User has an authenticated session to the target org (check: OrgAuthenticated)

Operations Reference

Operations grouped by purpose. Use these as the building blocks for the workflows above.

Summary

Dependency graph

graph TD
  discover_data_space["discover-data-space (read)"]
  verify_stdm_availability["verify-stdm-availability (validate)"]
  discover_agents["discover-agents (read)"]
  list_available_models["list-available-models (read)"]
  find_sessions["find-sessions (read)"]
  get_conversation_details["get-conversation-details (read)"]
  get_llm_step_details["get-llm-step-details (read)"]
  obtain_sfap_token["obtain-sfap-token (validate)"]
  replay_turn_prompts["replay-turn-prompts (read)"]
  score_exact_matches["score-exact-matches (read)"]
  judge_function_args["judge-function-args (read)"]
  judge_text_responses["judge-text-responses (read)"]
  generate_csv_report["generate-csv-report (write)"]
  generate_html_report["generate-html-report (write)"]
  discover_data_space --> verify_stdm_availability
  verify_stdm_availability --> discover_agents
  verify_stdm_availability --> find_sessions
  discover_agents --> find_sessions
  find_sessions --> get_conversation_details
  get_conversation_details --> get_llm_step_details
  get_llm_step_details --> replay_turn_prompts
  obtain_sfap_token --> replay_turn_prompts
  replay_turn_prompts --> score_exact_matches
  score_exact_matches --> judge_function_args
  obtain_sfap_token --> judge_function_args
  replay_turn_prompts --> judge_text_responses
  obtain_sfap_token --> judge_text_responses
  score_exact_matches --> generate_csv_report
  judge_function_args --> generate_csv_report
  judge_text_responses --> generate_csv_report
  generate_csv_report --> generate_html_report

Read operations

discover-data-space

Discover active Data Cloud data spaces in the org

Call: getDataSpaces — status: implemented

discover-agents

Query STDM for distinct agent + version combinations that have session traces

Call: queryCdpAnsiSqlV2 — status: implemented

Depends on: verify-stdm-availability

list-available-models

List Einstein LLM models configured in the org to help the user choose a target model

Call: queryAiModel — status: implemented

find-sessions

Find recent Agentforce sessions in STDM filtered by agent API name, version, and date range

Call: queryCdpAnsiSqlV2 — status: implemented

Inputs:

• agent_api_name (String) — Agent API name selected in discover-agents
• agent_version_api_name (String) — Agent version API name selected in discover-agents
• start_date (String)
• end_date (String)
• max_sessions (Integer)

Depends on: verify-stdm-availability, discover-agents

get-conversation-details

Fetch interactions (turns) for each session from the STDM Interaction DMO

Call: queryCdpAnsiSqlV2 — status: proposed

Depends on: find-sessions

get-llm-step-details

Fetch interaction steps including LLM inputs and outputs for each interaction

Call: queryCdpAnsiSqlV2 — status: proposed

Depends on: get-conversation-details

replay-turn-prompts

Extract structured messages (system/assistant/user) from STDM LLM step inputs, convert to OpenAI-format messages array preserving role positions, extract tool definitions dynamically from gen_ai.request.tools, classify each step as router (topic) or action based on tool signatures, and replay via SFAP REST API POST /ai/gpt/v1/chat/generations with tool_choice=required

Status: proposed

Inputs:

• target_model (String) — Target model ID (llmgateway__ prefix, e.g. llmgateway__OpenAIGPT41)

Depends on: get-llm-step-details, obtain-sfap-token

score-exact-matches

Compute exact-match scores for topic agreement and function name accuracy

Status: proposed

Depends on: replay-turn-prompts

judge-function-args

LLM-as-judge scoring of function arguments when function names match (exact or lenient)

Status: proposed

Depends on: score-exact-matches, obtain-sfap-token

judge-text-responses

LLM-as-judge scoring of text responses for turns that end with text output

Status: proposed

Depends on: replay-turn-prompts, obtain-sfap-token

Write operations

generate-csv-report

Generate CSV report with per-turn replay results and aggregate summary scores

Status: proposed

Inputs:

• output_csv (String)

Depends on: score-exact-matches, judge-function-args, judge-text-responses

generate-html-report

Build a self-contained HTML analysis report with verdict badges, scorecards, and failure breakdown

Status: proposed

Inputs:

• html_output (String)

Depends on: generate-csv-report

Other operations

verify-stdm-availability

Verify STDM (Session Trace Data Model) is queryable in Data Cloud

Call: queryCdpAnsiSqlV2 — status: implemented

Depends on: discover-data-space

obtain-sfap-token

Obtain an SFAP JWT token via org session authentication and the /ide/auth endpoint

Status: proposed

Inputs:

• org_alias (String) — Authenticated org alias for token resolution