Only three months after the debut of Agentforce 1.0, Salesforce launched Agentforce 2.0, the latest iteration of its autonomous AI platform which allows organizations to build and deploy autonomous agents for automated workflows. The updates include a new library of pre-built skills and workflow integrations for rapid customization, the ability to deploy Agentforce in Slack, and advancements in agentic reasoning and retrieval augmented generation (RAG).
Marc Benioff, Chair and CEO, Salesforce, billed Agentforce 2.0 as a “digital labor platform” intended to enable thousands of AI agents to collaborate with thousands of humans in jobs that combine human interactions and data-inquiry tasks – such as customer service agents, sales reps, marketers as well as many other types of desk/knowledge workers. Citing the global labor shortage (sources: Conference Board, HR Dive, McKinsey, U.S. News & World Report), Benioff argued that a digital labor force could help companies grow despite human labor constraints.
“I was talking to the CEO of a large bank that has already deployed part of Agentforce. He’s wanted to go into another geography, another huge market. He couldn't go because he didn't have the people. He had the systems, he had the brand, he had the trust, he had the regulatory capabilities, but he had not made the leap,” Benioff said. “When he [understood] that agents could deliver the labor in that market, he realized that this is a moment of abundance for his bank. He had been constrained by people. But when you have digital labor, you have a level of flexibility that you did not have before.”
Benioff emphasized Salesforce’s platform-based approach to rolling out agentic AI, as illustrated below, and that Agentforce has already been shipped to all 135,000 Salesforce customers. “They have it now. They just need to make the decision that they're ready to go to agents and turn it on,” he said.
In a press briefing, Sanjna Parulekar, VP, Product Marketing with Salesforce said that the best way to think about an AI agent is to start with the role or job that it will automate. The agent then needs access to data – unstructured and/or structured data. And if these agents are acting on an individual’s behalf (or a company’s) then the channels on which it ‘speaks’ need to be defined, as well as the guardrails required to constrain it.
“When you think about the actions that an agent can perform, it could be executing a process that you already have – a workflow, a prompt template or some combination of these or other things. When it comes to the skills and actions, the skills are really those predefined packages of these actions and instructions we give an agent,” Parulekar said. “A skill could be sales development or sales coaching or customer service and embedded in there are actions that help the agent perform that job. That could be delivering an answer to frequently asked questions, returning merchandise, something like that. Embedded in those skills as well are instructions which are absolutes: Always do this, never do that. Those types of statements help the agent function within a certain set of guardrails.”
MuleSoft is a platform that helps businesses connect systems, automate processes, and manage APIs. With the new MuleSoft for Flow, users can create low-code workflows that span systems with pre-built connectors for building multi-system workflows. This capability allows APIs to be turned into Agentforce actions and, via the new MuleSoft API Catalog, those APIs are made discoverable. The new MuleSoft Topic Center enables teams to describe APIs with Agentforce metadata so that these “points of connectivity” into other systems can be turned into an Agentforce skill or action.
MuleSoft for Flow, MuleSoft API Catalog, and Topic Center will be generally available in February 2025.
Agentforce 2.0 is now deployable in Slack and can be brought into any Slack conversation. Slack users can start a conversation directly from the Agentforce Hub, or @ mention Agentforce agents through DMs or in channels. The following animation shows a human interacting with the Agentforce “Deal Support Agent.”
Agent Builder now features pre-built Slack Actions such as “Create Canvas” or “Message Channel” that enable teams to enhance existing agents or create new ones. With Slack Enterprise Search, Agentforce can draw from public and permissioned information in Slack and thus improve the relevancy of responses and actions.
Agentforce in Slack, Slack Actions in Agent Builder, and Slack Enterprise Search will be generally available in January 2025.
In addition to the above, Agentforce 2.0 also includes the following new features:
The full release of Agentforce 2.0 will be generally available in February 2025, with specific features being released in advance beginning today (December 17, 2024). Natural language creation of agents in Agent Builder will be generally available in January 2025.
