What is Agentforce? Salesforce AI agents, explained

Leer este artículo en español →

Agentforce is Salesforce's platform for building, running and governing AI agents: software that understands a request, reasons about it and takes action inside your business processes, with your own data and under your own rules. If you searched for Agentforce and found mostly slogans, this is the plain version: where it comes from, what is inside an agent, how you build one, how it connects to other systems through MCP and what changes in a real process.

Joaquín País, CEO of ShowerThinking, also explains it on video:

What is Agentforce?

Salesforce describes Agentforce as a proactive, autonomous AI application that answers questions, takes actions and improves productivity. Put simply, an Agentforce agent does not stop at suggesting what to do. It does it, inside Salesforce and inside the systems you give it access to, and it leaves a trace of every step.

That is the difference with a classic chatbot. Salesforce sums it up on a slide from its Agentforce keynote at Dreamforce: most agent platforms only answer questions; Agentforce can take action.

From Einstein Copilot to Agentforce 360

Agentforce has a history, and it helps to know it, because the old names still show up in documentation and forums:

How does Agentforce work?

An Agentforce agent has four parts:

1. Atlas, the reasoning engine. When a request comes in, it works out what the person wants, breaks it into steps and decides what to do next. Salesforce calls it the Atlas Reasoning Engine.
2. Subagents. Each one handles a single job, such as scheduling a visit or answering a question from approved content. The agent picks the right one for each request.
3. Instructions. Each subagent has instructions in plain language that say how to behave and where the limits are.
4. Actions. What the agent can actually do. They are built from what you already have in Salesforce: flows, Apex code, prompt templates and APIs.

That is why a good agent looks more like a well-documented process than a clever chat: if the instructions are vague or the actions do not exist, reasoning will not fix it.

Data and trust: Data 360 and the Einstein Trust Layer

An agent is only as good as what it knows. Agentforce agents are grounded in your Salesforce data, in Customer 360 and Data 360, so they answer from your records and documents, not from what the model remembers about the world.

• Same permissions. The agent respects the permissions you already have in Salesforce: it sees what Salesforce lets it see.
• Einstein Trust Layer. Between the agent and the language model sits a layer that protects sensitive data, checks responses and keeps an audit trail.
• Data 360 underneath. According to Salesforce's own training, Data 360 must be enabled to use Agentforce, and features such as the Trust Layer and data libraries depend on it. Connecting your own sources to Data 360 is a separate step, needed when the agent has to work with more than the standard objects.

How do you build an agent in Agentforce?

You do not start from a blank page. Salesforce's own advice is that the fastest path to value starts with pre-built agents that already know your business:

1. Start from a pre-built agent. There are agents for common jobs, such as customer service, already connected to your customer data.
2. Customize it in Agentforce Builder. Describe the job in plain language, add subagents and actions, and test it in a preview conversation before anyone else sees it.
3. Lock down what cannot be improvised with Agent Script. For processes where improvisation is not acceptable, Agent Script mixes the model's flexible reasoning with fixed steps that run the same way every time. Salesforce calls it dynamic reasoning with deterministic control.

Agentforce and MCP

An agent also needs to reach systems outside Salesforce, and that is where MCP comes in. The Model Context Protocol is an open standard for how AI models connect to external tools, systems and data, and Agentforce supports it natively:

1. An admin adds an MCP server, often a verified one from the AgentExchange catalog.
2. They choose which of its tools the agent may use.
3. They assign those tools to a subagent as actions.

No custom integration for every system. It also works the other way round: with AIforce, Salesforce data and permissions reach assistants such as Claude and Slack. We cover it in What is Salesforce AIforce?

The work starts after launch

This is the part most teams underestimate. Traditional software follows a predictable lifecycle: you build it, test it and deploy it, and tomorrow it behaves the same way. An agent does not, because what it does depends on the conversations it gets. So Agentforce comes with tools for the work after launch:

• Testing against many conversations before going live.
• Observability to trace every session step by step and see why the agent did what it did.
• Agent Optimizer, which finds where conversations go wrong and suggests fixes.

In Salesforce's words, with agents the outcome is what drives the build process.

What changes in a real process: an example from pharma

A doctor's office asks to move a visit with the company's medical rep.

Without an agent, the message waits in an inbox until someone checks the rep's calendar, proposes new times and updates the account by hand.

With Agentforce, an agent reads the request, checks the calendar, offers two slots, books the visit and logs it on the account, within the rules you set. If the doctor mentions a side effect, the agent does not answer it: it hands the case to the pharmacovigilance team with the context already captured. Afterwards, the session shows every step the agent took.

Before you start with Agentforce

1. One process with a clear outcome. Not ten at once.
2. Clean data and permissions. The agent will use them exactly as they are, including the permissions nobody has reviewed in years.
3. A plan for after launch. Someone has to watch the sessions and tune the agent, because it only improves when someone is watching.

The second point delays more projects than any other: an agent with the wrong permissions either sees too much or cannot see what it needs.

Agentforce FAQ

What is Salesforce Agentforce?

It is Salesforce's platform for building and governing AI agents that take action inside business processes, using Salesforce data and permissions.

What happened to Einstein Copilot?

Agentforce replaced it. A copilot helps a person who still does the work; an agent takes action. The whole platform is now called Agentforce 360.

How does an Agentforce agent work?

The Atlas Reasoning Engine understands the request, breaks it into steps and chooses what to do. The agent has subagents, one per job, each with plain-language instructions and actions: flows, Apex, prompt templates and APIs.

Does Agentforce require Data 360 (formerly Data Cloud)?

Yes, Data 360 must be enabled to use Agentforce. Connecting your own data sources to Data 360 is a separate step, needed when the agent has to work with data beyond the standard objects or with documents.

What are topics and subagents in Agentforce?

The same thing under a new name. Each subagent, formerly a topic, covers one job of the agent, with its description, instructions and actions.

What is Agent Script?

It is the way to fix steps that must always run the same way inside an agent, combining the model's reasoning with deterministic logic.

Does Agentforce support MCP?

Yes. Agentforce can use MCP servers as actions in a subagent, with control over which tools from each server the agent may use.

Sources

• Salesforce, Agentforce
• Salesforce Developers, Agent Script Reference: topics renamed subagents, April 2026
• Salesforce, Agentforce 360 announcement, Dreamforce 2025
• CX Today, from Einstein Copilot to Agentforce
• Trailhead, Data Cloud and Agentforce
• Salesforce, "From Fast Start to Real ROI", Agentforce keynote, Dreamforce 2026
• Salesforce, "How Agentforce MCP Support Connects Agents to Your Tools with Trust"