Salesforce Expands Agentforce With Seven Named AI Agents and Enterprise Controls

Salesforce’s main pre-Dreamforce message was not simply that it has more AI assistants. It is packaging Agentforce as an enterprise execution layer: named agents designed for defined business jobs, grounded in Customer 360 data and paired with a proposed governance architecture for managing AI at scale. The seven-agent portfolio was announced September 11, while related partnerships and deployments were highlighted as Dreamforce opened on September 15. 749

The seven Agentforce agents and their jobs

Salesforce introduced seven “job-ready” agents spanning sales, service, commerce, employee support, supply chain, and customer experience. The agents are intended to use Customer 360 data along with prebuilt skills, actions, and data models rather than requiring customers to construct every workflow from scratch. 27

• Casey is a customer-service agent for resolving issues across voice, SMS, WhatsApp, and web chat, including FAQs, returns, account management, and escalation to a human when needed. 7
• Paige handles IT and HR requests. 29
• Carter supports shopping journeys, including product discovery, comparison, questions, and in-chat checkout. 7
• Piper qualifies inbound sales pipeline. 28
• Marshall supports supply-chain and back-office processes. 29
• Fin is focused on complex customer-experience workflows. 15
• Hunter is designed for outbound sales. 78

Salesforce said six of the agents were generally available. Hunter was in pilot, with general availability planned for November 2026. 812

Why Hunter is the technical centerpiece

Hunter is the first application Salesforce associated with a new long-horizon Agentforce runtime. Rather than completing work within one interaction, the runtime is designed to pursue a goal over days or weeks. Salesforce described three underlying capabilities:

• Memory, so an agent can retain relevant working context over time;
• Durable execution, so work can continue after pauses or failures; and
• Dynamic steering, so people or systems can redirect the agent while it is operating. 512

That framing matters because it shifts the product promise from answering prompts to advancing a continuing business objective—for example, outbound prospecting—within enterprise permissions and controls. The operational reliability of that promise will depend on real-world deployment results, which were not independently established by the announcement materials.

The scale claim: 7 billion Agentic Work Units

Salesforce said it had worked with customers on thousands of agentic deployments and delivered 7 billion Agentic Work Units (AWUs) across Agentforce and Slack in the preceding two years, including 3.2 billion in Q2. 67

The figures show the scale Salesforce says its platforms have processed, but they should not be read as independently audited evidence of productivity gains, cost savings, or return on investment. An AWU is an operational metric defined by Salesforce; enterprises evaluating Agentforce still need task-level measures such as accuracy, escalation rates, cycle time, customer outcomes, and the cost of human oversight. 14

The governance pitch: Trusted Enterprise AI Harness

Alongside the agents, Salesforce introduced the Trusted Enterprise AI Harness, a common and composable architecture intended to help companies provide agents with business context while applying enterprise controls. It brings together six areas Salesforce calls context, agency, action, governance, security, and models. 3641

A central component is the AI Control Plane, which Salesforce positions as a way to oversee agents and models across an open, composable AI environment. The aim is to avoid rebuilding context, policy, security, and management capabilities for every separate agent or AI experience. 36

This is the strategic connective tissue behind the Agentforce announcements: Salesforce is competing not only on individual agents, but on its ability to make those agents usable with existing customer relationships, processes, permissions, and data controls.

What Dreamforce added to the story

Dreamforce ran September 15–17 at San Francisco’s Moscone Center, with more than 43,000 attendees expected. Salesforce scheduled a conversation between Chair and CEO Marc Benioff and OpenAI CEO Sam Altman about advancing the AI frontier and business applications. 2228

On the opening day, Salesforce also pointed to deployments and partnerships that extend Agentforce beyond a standalone CRM feature:

• Adecco Group said it would scale Agentforce Coworker to more than 40 countries and roughly 27,000 employees. Salesforce described the deployment as embedded in its platform and built on Anthropic’s Claude. 1748
• Google Cloud and Salesforce expanded their partnership to support agent reasoning and actions across shared infrastructure and data, while bringing Salesforce workloads to Google Cloud through Hyperforce. 49
• Siemens said its Teamcenter technology would work with Agentforce to provide engineering-grade answers in sales, service, and customer workflows. Salesforce also said Agentforce would qualify inbound leads for Siemens’ 18,000 sellers. 52

The takeaway for enterprise teams

The announcement positions Agentforce as more than a chatbot catalog. Salesforce is trying to combine packaged, role-specific automation with CRM data, longer-running agent execution, and centralized controls for a wider ecosystem of models and systems.

For buyers, the practical question is not whether an agent has a memorable name. It is whether a defined workflow has the right data access, permissions, escalation paths, observability, and measurable business outcome. Salesforce’s new agents and AI Harness directly address those requirements in product positioning; the remaining test is how consistently they perform in production.