The true value of Salesforce Agentforce isn’t just launching an AI chatbot on a website — it’s channel abstraction.
Instead of building siloed bots for Web Chat, Email, and Voice, modern architecture decouples reasoning from presentation. A single Agentforce instance — configured with unified Topics, Actions, Guardrails, and Data 360 retrievers — serves as a central intelligence engine. The underlying delivery pipes and runtime adapters modify how that core intelligence presents itself based on where the user is interacting.
The Adaptive Runtime: How One Agent Serves Three Modalities
When an inbound event triggers an Agentforce topic, the Atlas Reasoning Engine executes the exact same underlying Flow or API action. However, the connection key determines the output adaptation:
- Web & Live Messaging (Synchronous Digital): Operates over persistent WebSocket connections via Messaging for In-App and Web (MIAW). The response adapter injects rich UI components — such as quick-reply buttons, interactive carousels, or dynamic forms — directly into the web client.
- Email & Async Pipelines (Asynchronous Text): Triggered via Email-to-Case or API ingestion. The exact same reasoning chain generates formal, multi-paragraph markdown or HTML responses, taking into account long thread histories and attached documentation.
- Voice & Telephony (Real-Time Audio): Integrated via Service Cloud Voice or Amazon Connect. The runtime formats responses into short, conversational turns optimized for low-latency Text-to-Speech (TTS) synthesis and live interruption/barge-in handling.
Front-End Choices for Internal CRM Users
For internal enterprise teams working within Salesforce, surfacing this central AI engine happens through three distinct front-end patterns:
1. Agentforce Coworker
The Ubiquitous Workspace Surface
Agentforce Coworker transforms the standard Global Search bar into a persistent conversational workplace assistant.
- Interaction Pattern: Ubiquitous and conversational. Users switch between standard search andAsk Mode at the top of the application.
- Scope & Grounding: Pulls plain-language answers across CRM records, connected Data 360 objects, and Slack threads.
- Capabilities: Executes natural language record updates directly from the search bar (with permission checks) and acts as a master router to dispatch tasks to specialized backend agents.
- Best Used For: Rapid internal lookups, ad-hoc record mutations, and cross-application context discovery.
2. Sidebar Panel (Agentforce Assistant)
The Contextual Record Co-Pilot
While Coworker is workspace-wide, the Sidebar Panel is tied directly to the record or conversation currently open on the user’s screen.
- Interaction Pattern: In-context and passive/assistive. It monitors the active record (Case, Account, Opportunity, or Messaging Session) and provides real-time recommendations.
- Capabilities: Generates instant summaries of long email threads or chat histories upon page load, pre-fills outbound replies for human editing, and executes inline actions without navigating away from the page.
- Best Used For: Dedicated service workspaces, complex case investigations, and human-in-the-loop review steps during live customer interactions.
3. Custom UI (Custom LWC & Headless API)
The Bespoke Enterprise Workstation
For enterprise teams needing custom layouts, Lightning Web Components (LWCs) or headless API integrations decouple the Atlas reasoning engine from standard out-of-the-box Salesforce containers.
- Interaction Pattern: Tailored and embedded. Built as custom LWCs inside the Service Console or integrated into third-party desktops via APIs.
- Capabilities: Renders specialized UI cards, multi-system telemetry views, and custom action buttons driven by Agentforce outputs, while integrating directly into third-party CTI softphones or external agent frames.
- Best Used For: Complex contact center desktops, industry-specific workstations (such as insurance claims or clinical health views), and multi-screen agent environments.
Handoff Mechanics: Transitioning from Agent to Human
When an agent hits a guardrail, low confidence score, or explicit transfer request, the handoff behavior adapts based on the active channel:
- Chat Handoff (Instantaneous): The Agentforce engine invokes an Omni-Channel Flow, passing MessagingSessionId, ContactId, and EscalationReason. The chat session remains connected seamlessly while Omni-Channel routes the work item to an agent'sEnhanced Conversation Canvas , pre-populating a live AI conversation summary.
- Voice Handoff (Real-Time SIP Bridge): The telephony gateway s automated audio processing and routes the VoiceCall record via Omni-Channel. The SIP audio stream reroutes to the human agent's softphone headset, while the workspace automatically opens the real-time audio transcript player and a 3-bullet briefing panel.
- Email Handoff (Asynchronous Review Queue): For low-confidence or sensitive topics, the agent does not send the email. Instead, it assigns the Case to an Omni-Channel queue with a status of Pending Agent Review. When opened, the human agent finds theEmail Composer pre-populated with an AI-generated draft ready for review and editing.
Experience Site Deployment: Enhanced Messaging for Web
When surfacing Agentforce on Salesforce Experience Cloud (LWR or Aura sites), deployment relies on Messaging for In-App and Web (MIAW) — often referred to in the architecture as Enhanced Web Messaging or Enhanced Chat v2.
Unlike legacy Live Agent, MIAW operates on persistent asynchronous infrastructure (built on Hyperforce/Salesforce Edge), enabling persistent chat sessions, pre-chat context binding, and native token-based user verification.
Authentication & Context Binding Options
When embedding the Embedded Messaging LWC component on an Experience Site, you have two primary setup choices for identity:
- Authenticated (Credential / Token-Based Verification): When logged-in users open the chat, the web client passes a JWT/User token to Salesforce. The platform binds the active $User.Id, ContactId, and AccountId directly to the MessagingSession record before the Agentforce agent runs its first reasoning cycle.
- Guest / Unauthenticated Mode: For public-facing sites, the agent initiates in guest mode. Context must be passed either via a Pre-Chat form component (mapping fields like Name, Email, and Subject) or through custom JavaScript APIs (embedded_svc.bootstrap.settings) before invoking the agent.
Step-by-Step Execution Flow: Chat Handoff to Human Agent
The transition from an autonomous Agentforce instance to a human service rep in the Service Console follows a single, uninterrupted session thread. The customer’s chat window never reloads, ensuring zero friction.
1.Handoff Trigger Identification: Step 1: Intent or Guardrail Breach.
The customer requests a representative, or Atlas reasoning hits a system guardrail (e.g., consecutive low-confidence responses, negative sentiment threshold, or explicit action failure).
2.Omni-Channel Flow Invocation: Step 2: Routing Execution.
Agentforce executes a system transfer action that passes execution control directly to an Omni-Channel Flow.
The agent passes key context variables into the flow:
- MessagingSessionId (the active persistent session)
- ContactId & AccountId (resolved identity context)
- EscalationReason & AgentforceSessionSummary
3.Queue & Skill Evaluation: Step 3: Work Distribution.
The Omni-Channel Flow evaluates queue availability, required agent skills (e.g., Spanish Language, Tier-2 Billing), and real-time human agent capacity.
While waiting in the queue, the client UI displays a status message (“Connecting you to a representative…”). 4.Console Accept & State Sync: Step 4: Human Workspace Launch.
The work item pops up in the target representative’s Omni-Channel utility widget.
When the rep accepts the request:
- The Enhanced Conversation Canvas opens automatically inside the Service Console.
- The complete turn-by-turn transcript between the end-user and Agentforce renders seamlessly.
- An AI-generated Session Summary Card provides the rep with the customer’s intent and reason for transfer before their first greeting.
Technical Routing Architecture Summary
- Front-End Component: Embedded Messaging LWC placed on the Experience Site template footer.
- Channel Definition: Messaging for In-App and Web (MIAW) Channel.
- Primary Entry Point: Omni-Channel Flow configured with Route To: Agentforce Service Agent.
- Fallback & Escalation: Omni-Channel Flow reroutes the active MessagingSession record from the Agentforce Agent ID to a Human Queue/Skill set.
Delivery Integrity Across the Boundary
This handoff flow demonstrates why treating the deployment as a single agent instance matters. The end user on the Experience Site experiences a single continuous chat thread, while the internal service representative receives a fully populated workspace containing identity context, conversation state, and transfer reasoning.
By grounding the Experience Site deployment in MIAW and Omni-Channel Flows, the channel runtime handles the transport, identity passing, and UI rendering natively. The Agentforce agent remains focused purely on topic evaluation, action execution, and knowledgable handoff triggers.
Here is a consolidated structural overview of how all the components we have covered fit together for your full article:
Introduction & Single-Agent Core: Establishing the architecture where one Agentforce instance (Topics, Actions, Data 360) connects to multiple channels.
Channel Modality Adaptations: How the same agent execution logic formats outputs differently for Web Messaging, Email, and Voice.
Internal CRM Front-End Options:
- Agentforce Coworker: Global search and workplace conversational interface across CRM, Data 360, and Slack.
- Sidebar Panel (Agentforce Assistant): Contextual record co-pilot in the utility bar for active cases, emails, or calls.
- Custom UI: Specialized Lightning Web Components and headless API integrations for custom workstations or CTI frames.
Experience Site & Web Deployment (MIAW):
- Authenticated JWT token passing vs. Guest pre-chat context binding.
- Step-by-step Omni-Channel Flow routing from agent execution to human escalation.
Cross-Channel Handoff Mechanics: Real-time chat transfer, SIP audio rerouting for Voice, and asynchronous review draft queues for Email.
One Brain, Every Touchpoint: Building Adaptive, Multi-Channel Agentforce Architectures was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.