The Building Blocks of Agents

Agent

Agents are trusted conversational AI assistants. They increase productivity and reduce workloads by automating routine tasks and assisting with complex ones. They're more autonomous than other conversational AI solutions, so they can adapt to different situations, environments, and information. Agentforce agents can independently identify opportunities for action, anticipate next steps, and initiate tasks within the use cases and guardrails you specify, with or without a human in the loop.

An agent can be created from scratch or from a template.

Subagents and Actions

An agent includes a library of subagents and actions, which are your agent's most important assets.

Actions are how an agent gets things done in Salesforce. An agent uses an action to get information or perform tasks. For example, if a user asks an agent for help with writing an email, the agent can execute an action that drafts and revises the email and grounds it in relevant Salesforce data. Salesforce provides some standard actions for common Salesforce tasks, and you can create custom actions specific to your business use cases.

A subagent is a category of actions related to a particular job to be done. For example, a subagent called Deal Management can contain agent actions that help a sales rep get up to speed on their day, find relevant opportunities and contacts, create to-do items, and log calls. Subagents contain actions, which are the tools available for the job, and instructions, which tell the agent how to make decisions.

Data

Agents are only as useful as the data they're grounded in. Because Agentforce agents are built on the Salesforce Platform, they're integrated with your CRM data and you control what data they can access. You can ground your agent in other data sources, including knowledge articles and fields, your file uploads, or web sources by using tools like Agentforce Data Libraries and the Search the Web standard agent action. For unstructured data sources, you can build more advanced data retrieval solutions with Retrieval Augmented Generation (RAG).

Connections and Channels

Channels are the messaging platforms, apps, and interfaces that you can deploy an agent to. They're how you send your agent out into the world, and they also determine how your agent escalates complex or sensitive conversations. Channels include text- and voice-based channels, as well as employee-facing and customer-facing channels.

When you deploy an agent to a channel, you also create and manage a connection for that channel. A connection includes adaptive response formats that help your agent structure responses and deliver multimedia content, such as images, buttons, links, and videos. It also includes settings and Omni-Channel flows that help route conversations to and from an agent. Connections help you scale agent development and reduce repetitive setup by letting you build an agent one time and easily add it to multiple channels in Agentforce Builder. They also help the agent make customer and employee experiences more dynamic, channel-specific, and consistent.

Reasoning Engine

The reasoning engine orchestrates how an agent reasons, takes action, and generates responses.

When an agent is triggered or when a user asks a question or makes a request, the reasoning engine works behind the scenes.

• Routes the conversation to relevant subagents, through subagent classification or deterministic transitions.
• Resolves subagent instructions to build a prompt, including replacing reference to resources with actual values, running inline actions, and evaluating and applying conditional statements.
• Reasons with the LLM to analyze the available information in the prompt and determine next steps.

The Atlas reasoning engine is a graph-based reasoning engine, which uses Agent Script to separate an agent's big-picture workflow from its conversational skills. The result is hybrid reasoning, which combines probabilistic, LLM-based reasoning with deterministic, rules-based execution in the same engine.

Large Language Model (LLM)

Agents harness the power of an LLM to reason, communicate with users, and take action in your org.

The reasoning engine calls the LLM at different times during a task or interaction with a user. The number and size of the LLM calls depends on the task and which subagents and actions are launched.