Salesforce Unveils 7 Agentforce AI Agents Ahead of Dreamforce

Salesforce’s pre-Dreamforce Agentforce announcement focused on making enterprise AI agents more deployable: seven prebuilt, role-specific agents, a runtime for work that can span days or weeks, and a common governance architecture. Salesforce says the agents are connected to Customer 360 context and existing business processes rather than operating as standalone chatbots. 23

The seven named Agentforce agents

The portfolio assigns each agent a defined business job. Salesforce said six were generally available at announcement, while Hunter remained in pilot. 23

• Casey — customer service: Handles service requests across voice, SMS, WhatsApp, and web chat. Its prebuilt flows cover FAQs, returns, account management, and escalation to a human agent. Salesforce cited Engine, where its help agent fully resolved 50% of chat inquiries. Generally available. 23

• Paige — IT and HR service: Addresses employee IT and HR requests through Slack, employee portals, and existing workplace tools. Salesforce cited Autism Queensland, where Paige resolved 70% of administrative requests. Generally available. 23

• Carter — commerce: Helps shoppers find and compare products, answers questions, and supports in-chat checkout. Salesforce cited Hibbett AI, which it said handled 90% of core shopper journeys following a six-week deployment. Generally available. 23

• Marshall — supply chain: Orchestrates back-office and supply-chain processes, automates manual work with deterministic execution, and keeps an auditable record of actions. Generally available. 23

• Piper — inbound B2B sales: Engages, qualifies, and converts inbound leads across websites and inboxes for sales and marketing teams. Salesforce cited Asana, where it said Piper generated four times the conversation volume, and said a typical deployment takes 45 days. Generally available. 23

• Fin — customer experience: Runs complex customer-experience workflows across channels, using Operator for customer operations and Fin Apex custom models for CX. Salesforce cited Anthropic, where it said Fin autonomously resolved 79% of conversations it received. Generally available. 23

• Hunter — outbound sales: Researches prospects and runs outreach alongside sales representatives over an extended sales cycle. Salesforce cited Perk, where it said Hunter built 60% of pipeline. Pilot at announcement; general availability planned for November 2026. 23

The customer-result figures above are Salesforce-cited examples, not independent benchmarks. They are useful indicators of the workflows the company is targeting, but buyers should validate performance against their own data, integrations, escalation requirements, and measurement standards.

A runtime built for work that outlasts a chat

Salesforce’s new long-horizon runtime is intended to let an agent work toward a goal across days or weeks instead of ending with a single interaction. Hunter is the first named agent on that runtime. 36

The company describes three core capabilities:

• Memory preserves relevant context and progress across sessions.
• Durable execution allows an agent’s plan to continue, resume, and recover over time.
• Dynamic steering lets users redirect the agent as circumstances or priorities change.

In Salesforce’s example, a seller could ask Hunter to rescue at-risk deals before quarter-end; the agent would translate that instruction into a measurable goal, create a plan, and work toward it over time. 36

That distinction matters for enterprise deployments. A useful agent for prospecting or service is not only one that produces a response; it must also retain appropriate context, respect changing instructions, and reliably handle work that pauses or spans multiple systems.

Salesforce’s reported 7 billion Agentic Work Units

Salesforce said it had delivered 7 billion Agentic Work Units (AWUs) across Agentforce and Slack during the prior two years, including 3.2 billion in Q2. 23

AWUs are Salesforce’s activity measure for work performed through the platform. The total signals the scale of usage Salesforce reports, but it should not be treated as a direct measure of customer ROI, accuracy, cost savings, or autonomous task completion without supporting outcome data.

The Trusted Enterprise AI Harness: one control layer for agents and models

Alongside the agents, Salesforce introduced the Trusted Enterprise AI Harness, a common architecture intended to give organizations shared controls instead of requiring each team to build context, security, and governance mechanisms separately for every AI experience. 40

Salesforce groups the harness into six capabilities:

• Trusted Context
• Trusted Agency
• Trusted Action
• Trusted Governance
• Trusted Security
• Trusted Models

Its proposed AI Control Plane is designed as a central place to manage agents and AI systems across the organization. Salesforce says the approach supports Salesforce and third-party models, agents, and systems, with model selection based on factors such as accuracy, performance, cost, and business needs. 40

For teams evaluating agents, this is the practical heart of the announcement. The named agents provide packaged starting points; the harness is Salesforce’s answer to the harder operational question of how to control identities, permissions, data access, actions, and model choices as deployments expand.

Dreamforce context: enterprise deployment meets the wider AI debate

Dreamforce 2026 was scheduled for September 15–17 in San Francisco, with more than 43,000 attendees expected at the Moscone Center. Marc Benioff and OpenAI CEO Sam Altman were scheduled to discuss how OpenAI is advancing the AI frontier and partnering with Salesforce for business tools. 5061

The event arrived amid a broader debate about the pace of advanced-AI development. Reuters reported that Altman told OpenAI employees the company was open to slowing development alongside other labs amid growing safety concerns. 47

Salesforce’s announcement does not resolve that industry-level debate. Its emphasis is narrower: putting agents to work within enterprise data, workflows, permissions, monitoring, and governance. For organizations considering Agentforce, the key test is therefore not just whether a named agent can perform a task, but whether it can do so reliably within the controls their business requires.