Salesforce has fundamentally shifted how businesses think about AI. With Agentforce, the question is no longer “Can AI assist my team?” — it’s “How much of the work can AI actually do on its own?”
This guide covers everything you need to know about Agentforce in 2026: what it is, how it works under the hood, the types of agents you can build, deployment options, and how your team can get started today.
Agentforce is Salesforce’s platform for building autonomous AI agents — software that can reason through a problem, decide what to do, take action, and respond, all without a human typing a prompt first.
This is a significant leap from traditional chatbots or even Einstein Copilot (the predecessor). A chatbot follows a script. A copilot waits for a human to ask something. An Agentforce agent can be triggered by an event — an incoming email, a new lead, a case update — and independently execute a multi-step workflow to resolution.
Three layers make this possible:
With Agentforce 3.0 (released June 2025) and Agentforce 360 (GA January 2026), the platform now includes a Command Center for full agent observability, Model Context Protocol (MCP) integration, Agent Script for deterministic control flows, and over 100 prebuilt industry actions.
If your org previously used Einstein Copilot, you’re already partway there. But the two are fundamentally different:
| Einstein Copilot (old) | Agentforce (2026) | |
|---|---|---|
| Trigger | Human types a prompt | Human prompt or automated event |
| Execution | Suggests — human approves | Acts autonomously, with optional human handoff |
| Actions | Limited, co-pilot style | Full Apex, Flow, API, Prompt Template support |
| Channels | Lightning Experience only | Lightning, Mobile, Slack, Experience Cloud, Web, WhatsApp, Email |
| Builder | Einstein Copilot Builder | New Agentforce Builder (GA February 2026) |
The Agentforce (Default) agent is the direct successor to Einstein Copilot for internal/employee use. Existing Copilot configurations can be migrated into Agentforce topics and actions
Agentforce supports two primary agent categories, each with distinct deployment surfaces:
Customer-facing. Handles support queries, case deflection, appointment booking, and FAQ resolution — 24/7, without a human in the loop unless escalation is needed.
Deploys to: Experience Cloud portals, external websites, WhatsApp, Facebook Messenger, Email, Enhanced Chat
Internal-facing. Helps sales reps qualify leads, assess deal health, log activities, and surface insights — directly inside the tools they already use.
Deploys to: Salesforce Lightning Experience, Salesforce Mobile, Slack
Both types support natural language multi-turn conversations, human escalation/handoff, and full access to your Salesforce data via SOQL-respecting Apex actions.
Before building, every developer and admin should understand the key components:
A topic is an intent category — a named bucket that defines when the agent should engage. Example: “Lead Qualification” or “Order Management.” Each topic has a classification description (used by the Atlas engine to route user messages) and a scope (what the topic covers).
The quality of your classification description directly determines routing accuracy. Write it as if you’re explaining to a colleague when they should handle this type of request.
Prompt-level rules inside each topic. These guide how the agent behaves — tone, what to check first, how to handle edge cases, whether to confirm before acting.
The tools the agent uses to get work done. There are four types:
@InvocableMethod (most powerful for complex logic)Upload knowledge articles, PDFs, and policy documents. The ADL indexes them for RAG — so your agent answers from your content, not generic LLM training data.
System-level gatekeepers that include or exclude topics/actions based on conditions — deterministic logic that doesn’t rely on the LLM. Think of them as dynamic visibility rules for your agent’s reasoning.
The new Agentforce Builder (GA February 2026) simplifies the process significantly. Here’s the current setup flow:
Setup → Einstein Setup → Turn on Einstein
Setup → Agents → Turn on Agentforce toggle
Setup → Salesforce Go → Search "Agentforce (Default)" → Get Started → Turn OnIn Agent Builder → New Topic. Give it:
For Apex actions, annotate your class with @InvocableMethod:
@InvocableMethod(
label='Score Lead'
description='Returns a qualification score and summary for a Lead based on industry, company size, and activity history.'
)
public static List<Response> scoreLead(List<Request> requests) {
// your logic here
}After deploying, register in Setup → Agent Actions → New and select your Apex method. Tick “Show in Conversation” on at least one output — this surfaces results to the user and keeps them in agent memory for follow-up turns.
Use the Conversation Preview panel to simulate real interactions. You can see exactly which topic was selected, which action fired, and the input/output of each call — invaluable for debugging.
If your team has Salesforce developers, Apex actions are where Agentforce becomes genuinely powerful. Unlike standard actions, Apex lets you:
A few best practices for production-quality actions:
label and description — vague descriptions lead to wrong action selection.@InvocableVariable descriptions — they help the engine auto-map conversation context to input fields.Salesforce has evolved its pricing model alongside the platform:
For accurate pricing based on your org edition and expected volume, we recommend a direct conversation with your Salesforce AE — or reach out to us for an independent assessment.
Agentforce is genuinely exciting — but getting it right requires more than toggling a few settings. The biggest risks we see in DIY implementations:
At Tenetizer Technologies, we’ve helped Salesforce customers across Service Cloud, Sales Cloud, and Experience Cloud build production-ready Agentforce implementations. Our approach:
Tenetizer Technologies is a Salesforce implementation partner specialising in AI, automation, and custom development. Based in India and USA, serving clients globally.