Kwitko Coffee — a full multi-cloud Agentforce reference org

An end-to-end Salesforce build that wires one coffee business across D2C/B2B Commerce, Service Cloud, Field Service, Revenue Lifecycle Management (RLM), Data Cloud (+ the web browsing SDK), Salesforce Payments, and Agentforce — with 12 AI agents and a layer of deterministic + generative AI on top. Two brands (Kwitko Coffee on WooCommerce + Bean & Brew on a native LWR storefront) run as one unified customer profile.

It's also a reference + playbook: it ships the reusable Claude Code skills used to build it, a one-command deploy, and a headless MCP path to drive the agents.

Scale: 172 Apex classes · 12 Agentforce agents · 9 Flows · 8 LWC · 24 custom objects · 14 permission sets · Data Cloud (Identity Resolution + 8 Calculated Insights + Web SDK) · source API 66.0. Built DX-first. Most of it deploys from metadata; the genuinely-UI/Salesforce-gated steps are listed explicitly in Post-deploy.

---

What each cloud / capability includes

🛒 Commerce — D2C + B2B (WooCommerce and native LWR storefront)

• WooCommerce integration (bi-directional): order pull/push, product sync, customers, coupons, webhooks, return receipts — WooOrderService, WooOrderPull, WooOrderActionService, WooProductSync, WooCustomerService, WooCouponService, WooWebhookResource, WooCartResource, WooReturnReceiptResource. HMAC-verified inbound webhooks; Named-Credential outbound.
• Cart & recovery: Cart__c/Cart_Item__c, CartRestoreService, CartLinkService, CartQueueService, abandoned-cart capture → lead (AbandonedCartService, Abandoned_Cart_to_Lead flow), ReplenishmentService (consumable re-order).
• Bean & Brew native storefront (LWR, second brand): custom homepage components bbHero, bbBrandStory, bbFeaturedCollection, bbValueProps, bbNewsletter, bbThemeLayout; CartExtension calculators BeanBrewTaxCartCalculator (tax) + BeanBrewPromotionsCalculator (spend-based promo); buyer provisioning (Bean_Brew_Commerce_Buyer permset).
• Promotions / coupons / pricing: CouponService, CouponAnalyticsService, Discount_Rule__mdt.

🎧 Service Cloud

• Case lifecycle: record types + support processes, queues, escalation, CaseService, CaseResolutionService, caseServicePanel LWC, transcript persistence (CaseTranscriptUtil, CaseMessagingSessionLinker).
• Order-service fix tools (real actions, not just answers): ReturnService, CancellationService, ExchangeService, ReshipService, AddressUpdateService, OrderModifyService, FailedPaymentService, StoreCreditService, PasswordResetService, OrderStatusService, TrackingService.
• Entitlements / SLA: Kwitko_Entitlement_Access, milestones; Knowledge grounding for the agent; Omni-Channel routing for web chat (Kwitko_Web_Chat_Routing flow + queues).
• Consent + identity gate: ConsentService, IdentityService, OTP verification (VerificationService, RequestVerificationCodeAction, VerifyCodeAction), LoginLinkService, Chat_Verification__c.

🔧 Field Service (FSL)

• Work orders → appointments, Maintenance Plans (preventive, asset-based), scheduling (Dispatcher Console / Classic), service territories/resources.
• On-site invocables: FieldServiceTechActions (incl. collect payment), FieldQuoteAction (field quotes), FieldServiceAutopilot, FieldVisitToolkit, service reports + digital signature.
• Field Service AI: FieldVisitAI — a grounded pre-visit brief + draft service-report notes via ConnectApi.EinsteinLLM, surfaced in the fsVisitAssistant mobile LWC (record action).
• Maintenance self-service: Fs* action classes power the Maintenance Concierge agents (coverage, book/reschedule/cancel visits).

💳 Payments

• Salesforce Payments for D2C checkout (Bean & Brew), field payments (FieldServiceTechActions.collectPayment → Payment record + WO Payment_Status, mock gateway), and pay-now retry links for failed commerce payments (FailedPaymentService). Field_Service_Payments permset.

🔁 Revenue Lifecycle Management (RLM / Revenue Cloud)

Built + proven on a headless Subscription Management org (no RCA builder/Place-Quote engine — the agents are the front door):

• Subscription lifecycle on Asset: provision / amend (seats) / renew / cancel — SubscriptionConciergeService + Sub* actions.
• AI quoting on the standard Quote object (no engine needed): DraftQuoteAction + QuoteDraftService.
• RLM ↔ Field Service coupling: AssetSubscriptionLifecycle trigger keeps MaintenancePlan + ServiceContract in lockstep on renew/cancel (AssetLifecycleCouplingService).
• Real-time subscription insights: SubscriptionInsightsService recomputes Account.Insights_Subscription_* synchronously on every Asset change (trigger-driven, not a nightly batch).
• See skills/salesforce-rlm/use-cases-and-agent-patterns.md → "engine-gated org playbook."

🧠 Data Cloud (unified profile + predictive)

• Unified customer profile via Identity Resolution (both brands → one Individual), DLOs/DMOs (dataSourceObjects, objectSourceTargetMaps), and 8 Calculated Insights (customer profile / category affinity / return risk / service risk / return-churn / shipment health / order patterns / web engagement).
• Augmentation back to CRM: DataCloudAugmentationService writes unified-profile signals (segment, LTV, churn, NBA) onto the Account so agents read them live.
• Predictive pipeline: an Einstein Studio churn model + ChurnScoreService (writes Account.Churn_*), CustomerInsightsService (RFM/value/taste), AtRiskCampaignBuilder (churn → campaign activation).

🌐 Data Cloud SDK — web browsing capture

• Web SDK / engagement capture of anonymous browsing → Web_Event__c → the Web_Event DMO; custom ingest endpoint EngagementRest (@RestResource); device→person identity stitch (IdentityStitchService, WebIdentityService) so anonymous activity resolves to the unified profile after sign-in. Engagement_Tracking permset.

🏷️ Multi-brand

One org, two brands as one profile — brand is a dimension, not an identity: MessagingSessionBrandStamp (chat → brand) + OrderBrandStamp (order → brand) so recommendations/segments are brand-aware.

---

The 12 Agentforce agents

Service Agents (customer-facing, MIAW web chat — need a runtime BotUser):

Agent	What it does
Kwitko_Concierge_Web / Kwitko_Concierge_Web_Live	The storefront concierge: shopping + personalized recommendations and full order service (status, tracking, returns, refunds, cancellations, cases) and field-service maintenance — all in one chat, identity-gated.
Maintenance_Concierge_Web	Customer self-service for Field Service maintenance (coverage, book/reschedule/cancel visits) over WorkOrder/ServiceAppointment/ServiceContract.

Employee agents (internal / headless — invoked via AgentInvoker / MCP):

Agent	What it does
Kwitko_Concierge	Concierge orchestrator (internal): greet, capture lead+consent, recognize returning customers, build buyer-aware recommendations.
Product_Advisor	The recommendation brain — profiles the buyer, picks products + quantities, resolves discounts deterministically (the engine other agents call).
Inside_Sales	Re-engages abandoned-cart leads with a personalized one-time recovery discount (consent-gated).
Post_Purchase_Growth	Analyzes a completed order, recommends an in-stock product, issues a one-time coupon, emails the customer; winback on return.
Lead_Copilot	Reviews a Nurture lead's unified-profile metrics (LTV, purchases, recency, cart) and recommends + applies the next-best conversion step.
Subscription_Concierge	NL front door to RLM subscriptions: view / provision / amend / renew / cancel on the Asset object.
Renewal_Retention	Works the subscription install base: renewals due + Data Cloud churn risk → next-best action (at-risk save / renew+upsell), renews, logs plays.
Quote_Copilot	Drafts standard Salesforce Quotes (Quote + QuoteLineItem) from natural language, at catalog price with an optional discount.
Maintenance_Concierge	Internal/headless field-service maintenance self-service (the employee-agent twin of the web one).

Agent actions are thin Apex @InvocableMethod wrappers; every agent shares one privacy-gated context envelope (AgentContextService / EmployeeAgentContextService) and an anti-hallucination instruction grammar (never claim an outcome unless the action returned success this turn).

The AI (beyond the conversational agents)

AI	What it does	How
Churn model	Predicts account churn risk → Account.Churn_Score__c/Churn_Risk_Tier__c; drives renewal/retention plays + at-risk campaigns.	Einstein Studio model + ChurnScoreService + AtRiskCampaignBuilder
Recommendation engine	Deterministic product/quantity/discount strategy the agents call (so the LLM never invents prices).	RecommendationStrategyService, ProductSuggestionService
Field Visit AI	Grounded pre-visit brief + draft service-report notes for technicians.	FieldVisitAI via ConnectApi.EinsteinLLM + fsVisitAssistant LWC
Profile insights	Per-account RFM / value / taste / subscription value + next renewal, written to Account.Insights_* (real-time for subscriptions).	CustomerInsightsService, SubscriptionInsightsService, DataCloudAugmentationService

---

Autonomous automation, governance & certification

This isn't just request/response agents — much of it runs proactively:

• Scheduled sweeps + record-triggered flows drive the business without a human asking: AbandonedCartSweep, CartRecoverySweep, PostPurchaseSweep, Chat_Nurture_Sweep, Lead_Copilot_Nurture_Sweep, RenewalRetentionScheduler, ReplenishmentService, ChurnScoreService, ServiceInteractionLogPull, plus flows Auto_Invoke_PostPurchase_Agent, Auto_Post_Purchase_Offer, Winback_On_Return, Order_Close_Lead_On_Purchase, Inside_Sales_Handoff_To_Nurture, Sync_Product_to_WooCommerce.
• Deterministic discount governance — a full RFM-tier × channel matrix in Discount_Rule__mdt (New / Occasional / Recurrent / Lapsed × abandoned-cart / chat / post-purchase). The agents resolve a discount from the matrix; they never invent one.
• Secrets via protected custom metadata (Kwitko_Secret__mdt / KwitkoSecrets) — no keys in code.
• Agent certification — multi-turn evaluation (aiEvaluationDefinitions) + a live certification matrix and evidence under docs/proofs/; design docs under docs/agents/.
• Service automation — Case assignmentRules + escalationRules, sharingRules, and a Field Service offline mobile briefcase (briefcaseDefinitions).
• Cross-agent memory & traceability — Customer_Journey__c + JourneyLogger/JourneyService record who/what/when across every agent so actions are auditable and agents share context.

---

Quick start — deploy to a fresh dev/scratch org (one command)

# 1) Authenticate to the target org (or create a scratch org with config/project-scratch-def.json)
sf org login web --alias myOrg --set-default

# 2) Ordered deploy + permission-set assignment + a printed post-deploy checklist
scripts/deploy/deploy-to-new-org.sh myOrg
#   scratch/dev orgs without full Apex coverage:  scripts/deploy/deploy-to-new-org.sh myOrg NoTestRun

The script is idempotent (re-runnable), deploys in dependency order (objects → Apex → integration → Omni-Channel → messaging → flows → Data Cloud → security → UI), assigns the core permission sets, and prints the manual steps that can't be deployed.

Org prerequisites (toggle BEFORE deploying — none are metadata): Person Accounts ENABLED, Agentforce/Einstein on, Data Cloud provisioned, Messaging for Web (Enhanced Messaging) on, Omni-Channel on, sf CLI authenticated. See the full list the script prints.

---

Use it — drive the agents headlessly (Apex / MCP / Flow)

Employee agents are conversational, not event-driven — fire them headlessly from Apex, a Flow, a scheduled job, or an MCP client.

From Apex / CLI (the in-org generateAiAgentResponse, no OAuth):

sf apex run --target-org myOrg \
  --apex 'System.debug(AgentInvoker.callAgent("Subscription_Concierge","I am dana.demo@example.com — what am I subscribed to?").agentResponse);'

Multi-turn — pass back the returned sessionId:

AgentInvoker.Response r1 = AgentInvoker.callAgent('Quote_Copilot','Quote Acme Roasters 20 of the Coffee Club at 10% off');
AgentInvoker.Response r2 = AgentInvoker.callAgent('Quote_Copilot','make it 30', r1.sessionId);

Via MCP (headless): tools/agent-mcp/ is a small MCP server exposing ask_kwitko_agent(agentApiName, message, sessionId?) over an Apex REST wrapper around AgentInvoker — drive any employee agent from any MCP client. Design + registration: docs/mcp-agent-consumer.md.

Customer-facing (Service Agents): reached through the deployed Messaging-for-Web channel (not generateAiAgentResponse). A storage-free SCRT/SSE smoke test lives in scripts/ci/web-chat-conversation-smoke.sh.

Note: generateAiAgentResponse works for employee agents; Service Agents are exercised through the messaging channel.

---

Prerequisites (enable in the target org BEFORE deploying)

These are org features the deploy depends on — none can be toggled by metadata:

• Person Accounts ENABLED (irreversible; Setup → Account Settings). The Account model + agent logic assume it.
• Agentforce / Einstein Generative AI enabled.
• Data Cloud provisioned (for DLO/DMO/Calculated-Insight metadata to import).
• Messaging for Web (Enhanced Messaging) enabled (for MessagingChannel/EmbeddedServiceConfig).
• Omni-Channel enabled (for Queue/QueueRoutingConfig + the routing flow).
• (Optional, per cloud) Field Service managed package, Revenue Cloud / Subscription Management, Salesforce Payments — enable the ones whose features you want.
• Salesforce CLI (sf) installed and authenticated; API version 66.0 compatible. Adequate data storage (DE orgs are tight — see gotchas).

---

Repository layout

force-app/main/default/   # all metadata (source of truth)
manifest/package.xml      # deployable manifest (wildcards; excludes non-deployable DC + agent bundles)
config/                   # project-scratch-def.json (scratch org shape)
scripts/
  deploy/deploy-to-new-org.sh   # ordered, idempotent deploy + permset assign + post-deploy checklist
  ci/                           # CI deploy/smoke/secret-scan scripts (GitHub Actions)
  apex/  soql/                  # one-off Apex / SOQL helpers
tools/agent-mcp/          # MCP server exposing ask_kwitko_agent (see docs/mcp-agent-consumer.md)
docs/                     # build guides, orchestration design, CI runbook, MCP design, use-cases & QA
skills/                   # reusable Claude Code skills (salesforce-rlm/agentforce/service/field-service/d2c-setup)

---

Deploy via the Salesforce MCP

If you use the connected salesforce MCP instead of (or alongside) the CLI:

• deploy_metadata — deploy a source dir or the manifest. Run once per ordered group (objects → apex → integration → omni → messaging → flows → data cloud → security → ui), or once with manifest/package.xml.
• run_soql_query — verification, e.g. SELECT DeveloperName, BotUserId FROM BotDefinition, SELECT COUNT() FROM UnifiedssotIndividual__dlm (unified profiles > 0).
• assign_permission_set — assign the core permsets (below).
• run_apex_test — run the Apex test suite.

Core permission sets: Kwitko_Integration, Kwitko_Messaging_Ops, Engagement_Tracking, Predictive_Scoring (+ per-cloud: Field_Service_Payments, Subscription_Insights_FLS, Bean_Brew_Commerce_Buyer, Kwitko_Entitlement_Access, Brand_Catalog_Access, …).

---

Post-deploy steps (NOT metadata-deployable)

A) Agentforce agents

Bundles are .forceignore'd / excluded from the manifest. Publish + activate each via sf agent (Agent Script) or Agent Builder:

sf agent validate authoring-bundle --api-name Subscription_Concierge --target-org myOrg
sf agent publish  authoring-bundle --api-name Subscription_Concierge --target-org myOrg
sf agent activate --api-name Subscription_Concierge --version 1 --target-org myOrg   # publish does NOT auto-activate

• Agent type is immutable after first publish — Service Agents stay Service Agents; employee agents stay employee.
• Service Agents need a runtime BotUser; employee agents run as their default_agent_user.
• ⚠️ Service Agent bundles carry a default_agent_user that is org-specific — set it to your org's Einstein Service Agent *.ext user before publishing.
• Verify an employee agent headlessly: sf apex run --apex 'System.debug(AgentInvoker.callAgent("Product_Advisor","hello").agentResponse);'

B) Data Cloud — CRM data streams (metadata first, UI fallback)

Deploy mktDataTranObjects + dataStreamDefinitions after dataSourceObjects. If a fresh org rejects a stream with "no MktDataTranObject named <X>_Home found", create it once: Data Cloud → Data Streams → New → Salesforce CRM → pick the source object → map to the matching *_Home DLO → Save & Run, then re-run the deploy.

C) Data Cloud — Identity Resolution + Calculated Insights (Connect REST API)

GOTCHA: Identity Resolution silently yields 0 unified profiles if the identity key sources from an empty field. The objectSourceTargetMaps here map a populated field — don't re-point them at an empty one. After the stream runs:

sf api request rest "/services/data/v62.0/ssot/identity-resolutions/<Ruleset>/actions/run-now" --method POST --body '{}' --target-org myOrg
for ci in Customer_Agent_Profile Customer_Category_Affinity Customer_Return_Risk Customer_Service_Risk \
          Return_Churn_By_Account Shipment_Delivery_Health Order_Patterns_by_Demographics; do
  sf api request rest "/services/data/v62.0/ssot/calculated-insights/${ci}__cio/actions/run" --method POST --body '{}' --target-org myOrg
done

D) Person Account default record type (UI-only)

Setup → Profiles → System Administrator → Object Settings → Account → Record Types → set "Person Account" as default (the system RT isn't in metadata).

E) Messaging for Web — user verification (Salesforce-side)

Authenticated chat via setIdentityToken needs the deployment's user-verification enabled (no metadata field) — open a Salesforce support case, then register your JWK/JWKS under Enhanced Chat User Verification. Until then, setIdentityToken is ignored and sessions are Guest.

F) Email deliverability (DNS) · G) Storage

Outbound agent email from an unowned-domain OWEA fails Gmail DMARC — add a sending subdomain with SPF/DKIM/DMARC. DE/small orgs are storage-tight: don't run sf agent test on a storage-tight org (AI-Evaluation records are a hard-to-delete hog); a MIAW POST /conversation → 412 usually means storage is full.

---

Verify the deploy

bash scripts/ci/salesforce-post-deploy-smoke.sh myOrg

Checks the web messaging channel/config, stuck MessagingSessions, the Service Agent's BotUser, and Data Cloud Calculated-Insight visibility + Account augmented fields.

---

Reusable skills (/skills)

Product-agnostic, DX-first playbooks used to build this:

Skill	Covers
salesforce-rlm	Revenue Cloud / RLM — catalog, pricing, Asset lifecycle, RLM↔Field-Service, the headless-vs-RCA fork, the engine-gated org playbook.
salesforce-agentforce	Agentforce agents (Agent Script), Data Cloud, MIAW chat, page-layouts/Dynamic-Forms, multi-turn agent testing.
salesforce-service	Service Cloud — cases, Knowledge, Omni-Channel, order-service flows.
salesforce-field-service	Field Service — enablement, managed package, work orders/appointments, scheduling, Maintenance Plans, payments, mobile.
salesforce-d2c-setup	D2C/B2B Commerce — storefront, catalog, checkout, payments/tax, promotions, OrderSummary.

Works with any agent, not just Claude. Codex, Antigravity, and Cursor have all converged on the same SKILL.md format.

• Claude Code — copy skills/* into ~/.claude/skills/.
• OpenAI Codex — copy agent-skills/* into ~/.agents/skills/ (auto-loaded by description match).
• Google Antigravity — copy agent-skills/* into a workspace skills/ directory.
• Cursor / any AGENTS.md tool — the repo-root AGENTS.md gives always-on context (deploy command, headless invocation, gotchas).

agent-skills/ is a portable copy of skills/ with Claude-specific references scrubbed; regenerate it with scripts/convert-skills.sh.

Further docs

• docs/SALESFORCE-BUILD-GUIDE.md — full build narrative · docs/BUILD_TUTORIAL.md — step-by-step tutorial
• docs/CHAT-AGENT-ORCHESTRATION-DESIGN.md — multi-agent orchestration design
• docs/USE-CASES-AND-QA.md — use cases + Q&A · docs/AGENT_CERTIFICATION_SUMMARY.md — agent test/cert summary
• docs/mcp-agent-consumer.md — MCP agent-consumer design · docs/CI_CD_RUNBOOK.md — CI/CD

CI/CD

GitHub Actions handle PR validation + staging/production deploys (docs/CI_CD_RUNBOOK.md). Secrets (SFDX auth URLs, JWT keys, WooCommerce keys) are never committed — see .gitignore + scripts/ci/secret-scan.sh.

License

MIT — see LICENSE. Reference build; not affiliated with or endorsed by Salesforce.