Agentforce
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 ยท 2,076 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 ยท 2,076 words
Add missing citations, update stale details, or suggest a clearer explanation.
Agentforce is an agentic AI platform from Salesforce that lets organizations build, deploy, and manage autonomous AI agents across customer service, sales, marketing, and commerce. Salesforce introduced it at the Dreamforce conference in September 2024 and made the first agents generally available on October 29, 2024 [1][2]. The platform runs on the Salesforce CRM, plans and acts through a component called the Atlas Reasoning Engine, and grounds its agents in company data through Salesforce Data Cloud [2]. Over the following year Salesforce shipped Agentforce 2.0, Agentforce 2dx, and Agentforce 3, then rebranded the whole stack as Agentforce 360 in October 2025 [3][4].
Agentforce marks Salesforce's bet that the next phase of enterprise AI is not a chat assistant sitting beside a human, but software that takes action on its own inside business workflows. Salesforce CEO Marc Benioff has framed this as a shift from copilots to a "digital labor" platform, where agents handle whole tasks end to end and escalate to people only when needed [3][5].
Agentforce sits as a layer on the Salesforce Platform that turns the company's existing CRM data, automation, and integrations into a foundation for autonomous agents [2]. An agent built on it can read the context of a customer interaction or an automated trigger, decide what to do next, and carry out the steps within limits the business sets. When a request falls outside its scope, the agent can hand off to a human [6].
Salesforce positions this as a third wave of AI. The first wave was predictive analytics, the second was generative copilots that draft and summarize, and Agentforce represents agents that reason and act [1]. The pitch rests on a simple argument. Salesforce already holds much of the customer data and process logic that an agent needs, so an agent grounded in that data can act with more relevant context than a general assistant bolted on from outside [2].
The agents draw on large language models for understanding and planning. Salesforce lets customers choose among foundation models rather than locking them into one. The Atlas Reasoning Engine can use models from OpenAI and Anthropic served through Amazon Bedrock, and Salesforce later added support for Google Gemini, alongside Salesforce's own models [5].
The core of the platform is the Atlas Reasoning Engine, which Salesforce describes as the brain inside Agentforce. It analyzes data, makes decisions, and completes tasks, and it does this by simulating how a person thinks and plans [2]. In practice it breaks a request into smaller steps, retrieves the data each step needs, evaluates possible courses of action, refines its approach as it goes, and then acts. Salesforce has said agents built on Atlas hit notably better accuracy and relevance on its internal Fortune 500 benchmarks compared with earlier approaches [2][7].
Grounding is what keeps the agents tied to reality. Agentforce surfaces the data an agent needs through Data Cloud, the Salesforce data layer later branded Data 360, so agents can act on live customer records without building custom integrations for each one [2]. Salesforce unifies an agent's data, content, prompts, workflow, and metadata, and protects it with the Einstein Trust Layer, a set of guardrails for data access, security, and policy compliance [2].
For building agents, Salesforce offers prebuilt templates and a low-code tool called Agent Builder, so teams can start from out-of-the-box patterns and customize topics, instructions, and actions [2]. The 2025 releases reworked this into a conversational Agentforce Builder, where teams describe an agent in natural language and then refine it in a document-style editor, a low-code canvas, or a pro-code view [4].
The table below lists the main building blocks.
| Component | Role |
|---|---|
| Atlas Reasoning Engine | Plans, decides, and executes multi-step tasks |
| Data Cloud / Data 360 | Grounds agents in unified company and customer data |
| Agent Builder | Low-code and conversational tool to create and configure agents |
| Einstein Trust Layer | Guardrails for data access, security, and compliance |
| Prebuilt agents | Ready-made agents for service, sales, and other functions |
The first agent Salesforce shipped at general availability was Agentforce Service Agent, alongside the Agent Builder tool [2].
Service is the flagship use case. A service agent runs around the clock across channels like voice, WhatsApp, Facebook Messenger, and websites, answering questions, resolving cases, managing orders, and troubleshooting, then escalating harder issues to human reps with the full conversation handed over [2]. With Agentforce 2.0 in December 2024, Salesforce added skills for sales development and sales coaching [3]. A sales development agent works the top of the funnel, answering product questions, handling objections, and booking meetings for human sellers, while a coaching agent role-plays pitches and negotiations [6]. Salesforce also offers agents for marketing, commerce, and general platform tasks [6].
Salesforce has extended the pattern beyond customer-facing work. In 2025 it announced Agentforce Operations, aimed at back-office processes like auditing, onboarding, and data entry, with specialized agents working across disconnected systems [6].
Salesforce moved fast after the Dreamforce debut. The platform reached general availability on October 29, 2024, just weeks after its announcement [1].
In December 2024 the company released Agentforce 2.0, which it called the first digital labor platform for enterprises. That release added a library of prebuilt skills, stronger orchestration for complex multi-step tasks, and the ability to deploy agents directly inside Slack [3]. At the TrailblazerDX conference in March 2025, Agentforce 2dx let agents run proactively in the background without a human prompt and added developer tooling and an Agentforce API [8]. In June 2025, Agentforce 3 introduced a Command Center for real-time visibility into agent behavior and added support for the Model Context Protocol for open interoperability with outside tools and data [9].
The milestones are summarized below.
| Release | Date | Headline additions |
|---|---|---|
| Agentforce (GA) | October 29, 2024 | Service Agent, Agent Builder, Atlas, Data Cloud grounding |
| Agentforce 2.0 | December 17, 2024 | Prebuilt skill library, better orchestration, agents in Slack |
| Agentforce 2dx | March 2025 | Proactive background agents, developer API |
| Agentforce 3 | June 23, 2025 | Command Center observability, Model Context Protocol support |
| Agentforce 360 | October 13, 2025 | Conversational Builder, Agent Script, Voice, rebrand |
On October 13, 2025, at Dreamforce, Salesforce announced Agentforce 360 and described an "agentic enterprise" in which humans and agents work in one trusted system [4]. The 360 brand pulls together four pieces. The Agentforce 360 Platform builds and runs the agents, Data 360 supplies shared context and memory, the Customer 360 apps hold the business logic, and Slack acts as the conversational interface [4]. New features in this release include the conversational Agentforce Builder, an Agent Script language that gives builders precise control over agent behavior, Agentforce Voice as a native voice layer for real-time spoken conversations, and Intelligent Context, which turns unstructured documents into agent-ready knowledge [4].
Salesforce launched Agentforce with consumption pricing rather than a flat per-seat fee. The first model charged 2 US dollars per conversation, with standard volume discounts, which the company pitched as a predictable way to scale a workforce on demand [2]. The same per-conversation rate applied to the sales skills added in Agentforce 2.0 [3].
That model drew confusion and pushback, because a conversation was hard to predict and price against. In 2025 Salesforce broadened the options. It added Flex Credits, a consumption model where each agent action costs about 0.10 US dollars, and it kept per-conversation and per-user licensing as alternatives [10][11]. The result is a menu of pricing models that customers can mix depending on how and where they deploy agents [11].
Salesforce reports steady uptake. By the Agentforce 360 launch it said more than 12,000 customers were using the platform, naming companies like 1-800Accountant, OpenTable, Reddit, and Engine [4]. Reddit reported a 46 percent case deflection rate after deploying Agentforce and cut its average resolution time by 84 percent, from 8.9 minutes to 1.4 minutes [12].
Salesforce also runs Agentforce on its own help site, a setup it calls being Customer Zero, and points to its own deployment as proof the model works at scale [4]. On the financial side, Salesforce told investors that Agentforce reached 1.2 billion US dollars in annual recurring revenue in the first quarter of fiscal 2027, up 205 percent year over year, with combined AI and data ARR of 3.4 billion dollars [13]. More than half of Agentforce and Data 360 bookings in the quarter came from existing customers, a sign of expanding use rather than just new logos [13][14].
Agentforce competes in a crowded enterprise agent market. Microsoft pushes agents through Copilot and its Copilot Studio, ServiceNow and Google both sell agentic products, and a wave of startups target the same workflows [4][15]. Salesforce's wager is that owning the CRM data, the apps, and Slack as the interface gives its agents better grounding and reach than rivals that have to reach into Salesforce from outside [4]. Microsoft, by contrast, leans on its grip over productivity software and Azure, and the rivalry between the two has turned pointed, with each casting doubt on the other's agent strategy.
The table contrasts a few of the main offerings.
| Vendor | Agent offering | Anchor |
|---|---|---|
| Salesforce | Agentforce 360 | CRM data, Data Cloud, Slack |
| Microsoft | Copilot, Copilot Studio | Microsoft 365, Azure |
| ServiceNow | AI Agents | IT and workflow platform |
| Agentspace and related | Cloud and Gemini models |
Agentforce is one of the clearest signals that a major software vendor sees autonomous agents, not just chat assistants, as the product to sell. Tying agents to CRM data and existing automation gives Salesforce a real grounding advantage, and the consumption pricing is an attempt to charge for outcomes and work done rather than seats [2][4].
The limits are just as real. Analysts caution that a large customer count does not prove deep usage, since many deployments may stay small or experimental [15]. Agents still depend on the quality of the data they are grounded in, and large language models can produce wrong or fabricated answers, which is why Salesforce wraps them in the Einstein Trust Layer and human escalation [2][6]. The shifting pricing in the platform's first year, from a single per-conversation rate to a mix of models, reflects how hard it is to price agent work in a way customers find predictable [10][11]. Whether agents can hold accuracy and trust at full production scale, across many tasks at once, is the open question that will decide how far the platform goes.