What Is AgentForce?

You’ve probably heard the name Agentforce in Salesforce news, sales calls, or your own org’s release notes. But most explanations are either full of hype or full of jargon. This guide fixes that.

Below, you’ll learn what Agentforce is in plain English, how it actually works behind the scenes, what it can and can’t do, what it costs, and how to tell whether your business is ready for it. No fluff, and no promises Salesforce itself doesn’t make.

What Is Agentforce? (Quick Answer)

Salesforce first showed Agentforce at Dreamforce in September 2024. Since then it has grown from a set of ready-made agents into a full platform, now often called Agentforce 360, that includes tools to build, test, launch, and monitor agents.

Agentforce in plain English

Think about hiring a new team member. You give them a job description, access to the systems they need, a list of things they’re allowed to do, and a manager to call when something is above their pay grade.

Agentforce works the same way. You give an agent:

• A job (for example, “help customers track and return orders”)
• Access to information (order records, return policy, customer history)
• Tools (start a return, issue a label, update the case)
• Rules (never refund over $500 without a person, never share another customer’s data)
• A handoff path (send tricky cases to a live agent with full context)

The difference is that this “team member” works around the clock, handles many conversations at once, and does the same thing the same way every time you’ve told it to.

How Does Agentforce Work?

Every Agentforce agent follows the same basic loop. It understands the request, gathers the facts, figures out a plan, takes action, and then either finishes the job or hands it to a person.

1. Understand the request. A customer types “Where’s my order?” or a business event happens, like a new lead coming in. The agent reads the request in natural language and figures out what the person actually wants.
2. Retrieve trusted data. The agent looks up the facts it needs from Salesforce records, Data 360, knowledge articles, and connected systems. This is called grounding, and it keeps answers tied to your real data instead of guesses.
3. Reason and plan. The Atlas Reasoning Engine decides which subagent should handle the job and which actions to use, in what order.
4. Take action. The agent runs the actions you’ve approved, such as a Salesforce Flow, Apex code, an API call, or a prompt template. For example, it can update a record, create a case, or send a confirmation.
5. Finish or hand off. If the job is done, the agent replies or closes the loop. If the request is outside its rules, it passes the conversation to a person along with everything that has happened so far, so the customer doesn’t have to repeat themselves.

Key Parts of Agentforce

Agentforce isn’t one feature. It’s a set of building blocks that work together. Here’s how they stack up. (For a deeper look at the platform layer, see our guide to the Agentforce Platform.)

Agentforce Builder and Agent Script

Agentforce Builder is where you create and edit agents. Admins can describe what they want in plain language or use a visual Canvas view. Behind the scenes, Salesforce uses Agent Script, a language that blends natural-language instructions with fixed logic like if/then rules, variables, and transitions. That mix lets you keep the flexibility of AI while locking down the steps that must always happen the same way.

Developers can also work in code using Agentforce DX, the Salesforce CLI, and a VS Code extension, so agents can move through normal release pipelines like the rest of your Salesforce work.

Subagents (formerly “topics”)

A subagent is one specific job the agent can do, along with its instructions, actions, and limits. A service agent might have one subagent for order status, one for returns, and one for billing questions. In April 2026, Salesforce renamed “topics” to “subagents.” Nothing about how they work changed, but you’ll see both terms in older articles and documentation.

Actions

Actions are what turn an agent from a talker into a doer. An action can run a Salesforce Flow, call Apex code, hit an API, use a prompt template, or reach outside systems through MuleSoft. Salesforce includes standard actions, and you can build custom ones. Connecting the right systems is often the biggest part of a project, which is why Agentforce integration planning matters so much.

Atlas Reasoning Engine

Atlas is the “brain.” It reads the request, picks the right subagent, decides which actions to use, and checks its plan before acting. Salesforce calls this hybrid reasoning: AI handles the parts that need judgment, while fixed rules handle the parts that must be predictable. We break this down further in the primary function of the reasoning engine in Agentforce.

Data 360 (formerly Data Cloud)

An agent is only as good as what it knows. Data 360 brings customer and business data together from Salesforce and outside sources, including unstructured content like PDFs. Salesforce renamed Data Cloud to Data 360 in October 2025. Not every agent needs Data 360 on day one, but many advanced use cases depend on it.

Einstein Trust Layer

The Trust Layer is Salesforce’s set of safety controls for generative AI. It grounds answers in your data, masks sensitive details before they reach an AI model, checks for harmful content, keeps an audit trail, and relies on zero-data-retention agreements with outside model providers. It lowers risk, but it doesn’t replace your own permissions, testing, and review.

Testing and Agentforce Observability

Launching an agent is the start, not the finish. Salesforce provides a Testing Center for running test conversations in bulk before launch, and Agentforce Observability for tracking how agents perform after launch, including escalation rates, deflection rates, user feedback, and abandoned sessions. Pair this with your own reporting through Agentforce Analytics.

Channels, including Agentforce Voice

Agents can work on your website, in messaging apps, over email, in Slack, inside Salesforce apps, and by phone through Agentforce Voice. The same agent logic can serve several channels, so customers get consistent answers wherever they reach you.

Agentforce vs. Chatbots, Copilots, and Salesforce Flow

This is where most people get confused, so here’s a side-by-side view.

A key point: Agentforce and Flow are partners, not rivals. In most real builds, the agent handles the conversation and decision, and then calls a Flow to do the actual work in a reliable way. If you’re weighing where AI fits versus standard automation, our AI and automation solutions team usually starts with that exact question.

You may also see older references to Einstein Copilot. Salesforce folded that assistant into the Agentforce family as part of its rebrand, so it’s now part of the same platform rather than a separate product.

What Can Agentforce Do? Use Cases by Team

The best use cases share three traits: lots of repeat requests, data the agent can reach, and clear rules for what “done” looks like. Here’s what that looks like across teams.

Want to go deeper by team? See how we approach Salesforce for customer service, Salesforce for lead management, and Agentforce in Marketing Cloud.

A Real-World Example, Step by Step

Let’s say a customer messages an online store at 11:40 p.m.: “My order hasn’t shown up. Can you check it?” Here’s what they see, and what the agent is doing behind the scenes.

The customer got help at midnight, the support team starts the morning with one less ticket, and every step is logged. That’s the practical value of Agentforce: not just answering, but finishing the job safely.

How Much Does Agentforce Cost?

There’s no single price. Salesforce offers several models, and the right one depends on who uses the agent (customers or employees) and how much. Here’s what Salesforce lists as of fall 2026.

Why the number of actions matters

With Flex Credits, you pay for each action an agent takes, not each conversation. A simple FAQ answer might use one or two actions. A full return with identity checks, lookups, and a new order might use eight or more. The chart below shows how that adds up.

Quick estimate formula: monthly conversations × average actions per conversation × $0.10 = rough monthly Flex Credit cost. For example, 3,000 conversations × 4 actions × $0.10 ≈ $1,200 per month. To compare that against the time your team saves, try our Salesforce ROI calculator.

Benefits of Agentforce

It finishes tasks, not just conversations

Because agents can take real actions, they can close out common requests from start to finish instead of just pointing people to a form or a phone number.

It’s available 24/7 on many channels

Customers get help on nights, weekends, and holidays, on the channel they already use. Busy seasons don’t require the same jump in hiring.

It works with what you already have

Agentforce runs on the Salesforce Platform, so it can reuse your existing data, Flows, Apex, permissions, and security model. You’re extending your org, not bolting on a separate tool.

Your people focus on harder work

When agents handle the routine requests, your team can spend more time on complex cases, relationships, and decisions that need human judgment.

You can watch and improve it

Built-in testing and observability tools show where agents succeed, where they escalate, and where conversations drop off, so you can fix weak spots with real data.

Limitations and Risks to Plan For

Agentforce is powerful, but it’s not magic. Being honest about the limits up front saves money and frustration later.

• Bad data means bad answers. Duplicate records, outdated knowledge articles, and missing fields will show up in what the agent says and does. Many teams need a data cleanup or migration before launch.
• It can’t fix an unclear process. If your team can’t explain the rules for a refund, the agent can’t follow them either. Write the process down first.
• AI can still make mistakes. Grounding and the Trust Layer reduce errors, but they don’t remove them. High-risk actions should require human approval.
• Permissions need real thought. Decide exactly what each agent can read, change, send, or trigger. Start narrow and widen access over time.
• Usage costs can surprise you. Long, multi-step conversations use more actions. Set usage alerts and review consumption every month.
• Not every process needs AI. If inputs are fixed and rules never change, a standard Flow is cheaper and more predictable.
• The product changes fast. Names, builders, and pricing have shifted several times since 2024. Plan for ongoing upkeep, not a one-time setup.

Common myths about Agentforce

Is Your Business Ready for Agentforce? A Quick Checklist

Before you spend money, run through this list. If you can check off at least six of these eight, you’re in a good spot to start a pilot.

• You can name one specific, high-volume task you want the agent to handle.
• You can describe what “done” looks like for that task in one or two sentences.
• The data the agent needs already lives in Salesforce or a system you can connect.
• That data is reasonably clean and current.
• Your team has written rules for exceptions, limits, and approvals.
• You know when the agent must hand off to a person, and who receives the handoff.
• You have a baseline metric to beat, such as average handle time or tickets per week.
• Someone on your team will own the agent after launch and review it monthly.

Checked fewer than six? That’s normal. It usually means the first step is process and data work, not AI. A short Salesforce consulting assessment can show you what to fix first.

How to Get Started With Agentforce: A 7-Step Plan

1. Pick one use case. Choose a task with high volume, low risk, and clear rules. Order status and appointment booking are classic starting points.
2. Set your success metric. Decide how you’ll measure success, such as resolution rate, response time, or hours saved per week.
3. Audit your data. List every record, field, and knowledge article the agent will need. Fix gaps before you build.
4. Define actions and permissions. Spell out what the agent can read, create, update, and send. Build or reuse the Flows and Apex it will call.
5. Set guardrails and handoffs. Write the rules for when the agent must stop, ask for approval, or pass to a person.
6. Test hard before launch. Use the Testing Center to run normal questions, odd phrasing, wrong information, and attempts to break the rules.
7. Launch small, then monitor. Start with one channel or a slice of traffic. Review escalations, failed conversations, feedback, and credit usage every week at first, then monthly.

Most teams get faster results with a partner who has done this before. Our Salesforce implementation services cover setup and launch, and our Salesforce integration services connect the outside systems your actions depend on. After launch, ongoing support and maintenance keeps agents tuned as Salesforce releases change.

How Agentforce Has Changed Since 2024

If you read about Agentforce a year ago, some of what you learned is now out of date. Here’s a quick timeline of the biggest changes.

• September 2024Salesforce introduces Agentforce at Dreamforce as its next step beyond chatbots and copilots.
• October 2024Agentforce becomes generally available, priced at $2 per conversation.
• December 2024Agentforce 2.0 adds more ready-made skills and deeper reasoning.
• May 2025Flex Credits launch, moving pricing toward a pay-per-action model.
• June 2025Agentforce 3 adds a command center for monitoring agents and support for the Model Context Protocol (MCP).
• October 2025The platform is branded Agentforce 360, and Data Cloud is renamed Data 360.
• April 2026Agent “topics” are renamed “subagents,” with no change in how they work.
• Mid-2026The new Agentforce Builder with Agent Script becomes the main place to create agents, replacing the legacy builder in Setup for new agents.

Frequently Asked Questions About Agentforce

Final Thoughts: Start With a Problem, Not the Tool

Agentforce is one of the biggest changes to Salesforce in years. It moves AI from “helping people type faster” to “doing real work safely.” But the companies that get value from it don’t start by asking, “Where can we add AI?” They start by asking, “Which task is costing us the most time, and is it ready for an agent?”

Pick one task. Clean the data behind it. Write the rules. Test hard. Then expand. That’s how Agentforce turns from a buzzword into a working part of your team.