Salesforce on September 11, 2026 released a portfolio of seven prebuilt AI agents for sales, service, commerce, employee support and back-office work, six of them generally available and one in pilot, together with a new runtime that allows an agent to hold a goal across days and weeks rather than a single conversation.

In Short

Salesforce packaged seven ready-made AI agents, each built for one job such as customer support, IT help desk, online shopping, outbound sales, supply chain, inbound lead qualification and customer experience workflows. Six can be bought and switched on immediately; the outbound sales agent, Hunter, stays in pilot until November 2026 and is the first to run on new plumbing that keeps an agent working toward an objective over weeks instead of ending when the chat window closes. For marketing and revenue teams, the practical shift is that pipeline generation, chat commerce and service deflection now arrive as configured products rather than as components to assemble.

Seven agents, six of them available immediately

The announcement, issued from San Francisco, names each agent and assigns it a function rather than a capability. Casey handles customer service resolution across voice, SMS, WhatsApp and web chat, shipping with prebuilt handling for frequently asked questions, returns, account management and escalation to a human. Paige covers IT and human resources requests through Slack, internal portals and existing employee tools. Carter works the shopper path, helping buyers find and compare products, answering questions and closing the transaction through in-chat checkout. Marshall sits in the back office, orchestrating end-to-end processes with what Salesforce describes as deterministic execution and an audit record of every action taken.

Piper works websites and inboxes to engage, qualify and convert inbound leads into pipeline for business-to-business sales and marketing teams. Fin resolves customer experience workflows across channels, running on a customer operations agent called Operator and on Fin Apex, a set of models the company says are custom-trained for customer experience work. Those six are generally available.

Hunter is the exception. Described as an outbound sales agent that works a pipeline from research through outreach and collaborates with sellers over weeks and months, it is in pilot now with general availability set for November 2026.

Every agent connects to Customer 360 and inherits the customer records and business processes already held there. Customers can rename an agent, and each operates inside the buyer's own business rules, permissions and security model. Behaviour can be specified through Agent Script, the open-source language Salesforce published for agent behaviour, which mixes model reasoning with deterministic rules so that certain decisions follow fixed logic rather than inference. That combination is the quiet part of the release: a system that reasons freely is hard to audit, and the back-office agent in particular is sold on its audit record.

The runtime underneath Hunter

The second half of the announcement concerns infrastructure rather than packaging. Salesforce built a long-horizon runtime for Agentforce, and Hunter is the first agent to run on it. The company says more agents in the portfolio will move onto the runtime over time, and that customers will eventually build long-horizon agents themselves.

Three capabilities sit underneath. Memory carries context and progress between sessions, so a plan survives the end of an interaction. Durable execution keeps that plan running and allows the agent to resume or correct course when circumstances change. Dynamic steering adjusts behaviour in response to an individual user's feedback and direction.

The worked example Salesforce gives is a seller asking Hunter to rescue at-risk deals before quarter end. The agent converts that instruction into a measurable goal, builds a plan, and determines which tasks to complete, which tools and context are required, and where the guardrails sit between acting autonomously and requesting seller approval. Nothing in the announcement specifies how those guardrails are configured, how approval thresholds are set, or what happens when a long-running plan conflicts with a change in the underlying record.

The distinction matters more than the vocabulary suggests. The OECD, in a 34-page working paper published in February 2026, separated a single AI agent that acts with some autonomy from agentic AI, meaning multiple coordinated agents pursuing complex objectives over extended periods with minimal supervision. Sustained goal pursuit over weeks is the second category, and it is the category where oversight, liability and specification questions have been least settled.

Two of these agents arrived by acquisition

Piper and Fin did not originate inside Salesforce. Piper is the inbound sales development product built by Qualified, the San Francisco company founded in 2018 by former Salesforce executives Kraig Swensrud and Sean Whiteley. According to Salesforce's quarterly filing with the Securities and Exchange Commission, the company acquired Qualified in April 2026 for consideration valued at approximately $1.2 billion, of which roughly $1.1 billion was cash, recording $954 million of goodwill and about $290 million of intangible assets.

Fin is Intercom. The same filing records that Salesforce entered into an agreement in June 2026 to acquire Intercom, Inc., listed under the name Fin. A separate filing by Hercules Capital, an Intercom lender that committed $250 million in March 2026, put the transaction at approximately $3.6 billion.

Presenting both as members of a single portfolio is a product decision rather than a technical statement. Neither the announcement nor the filings describe how deeply either system has been rebuilt on Salesforce's own stack, and the release attributes Fin's performance to its own models rather than to Agentforce reasoning.

The customer figures, and what they leave out

Salesforce published six deployment statistics. Engine resolves half of its chat inquiries through a help agent named Eva. Perk builds 60% of its sales pipeline through Hunter. Autism Queensland resolves 70% of administrative requests through Paige. Hibbett AI covers 90% of core shopper journeys and went live in six weeks. Asana's website agent, Piper, now drives four times the conversation volume, with Piper deployments averaging 45 days. Anthropic resolves 79% of the conversations Fin sees without human involvement.

All six are vendor-supplied, and none carries a denominator, a measurement window or a definition of resolution. A resolution rate depends entirely on which contacts enter the funnel: an agent that handles password resets and order status will post a higher figure than one exposed to billing disputes. Four times the conversation volume describes activity, not outcome. Sixty per cent of pipeline built by an agent says nothing about what share of that pipeline closes.

One figure also carries a naming discrepancy worth flagging. Salesforce's April 2026 material on Engine, covered when the company opened its platform to external coding agents through Headless 360, named the customer service agent Ava, built in 12 days and handling 50% of customer cases autonomously. The September 11 release names the same agent Eva and gives the same 50% figure. The release does not explain the change, and it is not clear whether the earlier spelling, the later one, or a rename inside Engine accounts for it.

Work units as the headline metric

Salesforce framed the release with volume rather than revenue. Over the past two years, it says, 7 billion Agentic Work Units have been delivered across Agentforce and Slack, including 3.2 billion in the second quarter alone.

An Agentic Work Unit is Salesforce's own metric, first disclosed at its fourth-quarter fiscal 2026 results in February 2026, and defined as one discrete task accomplished by an agent: a prompt processed, a reasoning chain completed, or a tool invoked. The company positioned it explicitly against token counts, arguing that tokens measure consumption rather than completed work. At that February disclosure, the cumulative figure stood at 2.4 billion units, with 771 million recorded in the fourth quarter, up 57% quarter over quarter.

Set against that baseline, the arithmetic is the story. Cumulative units moved from 2.4 billion to 7 billion in roughly two quarters, and a single quarter now accounts for 3.2 billion, more than four times the quarterly figure disclosed seven months earlier. The metric remains defined and counted by the vendor, with no external audit and no published breakdown by agent type in the September release, so it measures platform activity rather than customer outcome. A tool invocation that fails still counts as work performed.

The platform layer around the agents

Three additions accompany the portfolio. AI Skills inside Agentforce Coworker lets an employee teach the agent how to complete a task once, then reuse that method across the workforce and across interfaces; it is in pilot now with general availability in October 2026. Multi-Agent Orchestration routes work between specialised agents so that a job crossing roles, systems or stages of a customer journey is handled as one coordinated sequence, and it is generally available. Agent Optimizer assists teams through the agent lifecycle, covering construction and refinement of agents, subagents and actions, performance testing, and analysis of session traces to identify what to change; general availability is set for October 2026.

Multi-Agent Orchestration is the component with the widest implications, because coordination between agents is where interoperability questions surface. Routing inside one vendor's platform is a solved problem in a way that routing between vendors is not, and the announcement describes the former.

Salesforce closed the release with a standard disclaimer that it may reference services or features still in development and unreleased, and that customers are directed to base purchase decisions on currently available functionality. Three items in the announcement carry future dates.

Measured against Salesforce's own research

The most useful counterweight to a long-horizon runtime comes from Salesforce AI Research. Its CRMArena-Pro benchmark, published on June 10, 2025, found that leading language model agents succeeded in 58% of single-turn business tasks and 35% of multi-turn ones across 19 business tasks and 4,280 query instances. Workflow execution proved the most tractable skill, exceeding 83% in single-turn conditions, while confidentiality awareness was a consistent weakness across every model tested.

That study measured agents completing tasks inside a conversation. The runtime announced on September 11 extends the horizon to days and weeks, which multiplies the number of turns, tool calls and state transitions between instruction and outcome. Salesforce has not published an updated benchmark measuring long-horizon performance, and the release offers no error rate, no intervention rate and no figure for how often a plan is abandoned or corrected.

The wider evidence base is mixed in the same direction. Google Cloud's survey of 3,466 senior business leaders found 88% of early adopters reporting positive return on agent deployments, a figure drawn from self-assessment rather than audited accounts. Marc Benioff said in July 2025 that agents were performing 30% to 50% of work inside Salesforce and resolving 85% of customer service inquiries, and the company has used its own operations as a proof site since.

What this changes for marketing and revenue teams

Two of the seven agents sit directly in marketing workflows. Piper occupies the inbound path, engaging traffic on a website and in an inbox, qualifying and converting it into pipeline. Carter occupies the commerce path, running product discovery and comparison and completing the purchase inside the conversation.

Carter is the more structurally interesting of the two, because in-chat checkout moves the conversion event off the product page and into a dialogue. That direction is not specific to Salesforce. Google set out the Universal Commerce Protocol at the National Retail Federation conference in January 2026, defining how agents discover a merchant's catalogue, build carts and complete payment; Salesforce joined its Tech Council in April 2026 alongside Amazon, Meta, Microsoft and Stripe. Adoption of that standard has lagged its endorsement list, with a scan in May 2026 finding 26 public sites carrying the required files out of more than three million checked. A conversion completed inside an agent conversation does not fire the same events as a checkout completed on a page, which leaves attribution and measurement to be rebuilt rather than reconfigured.

For service and support functions, the packaging argument is simple: prebuilt agents shorten the distance between purchase and deployment, and Salesforce cites deployment windows of six weeks for Hibbett and an average of 45 days for Piper. Those windows describe implementation, not payback.

The governance question is the one the release addresses least. Across the wider agentic advertising market, guardrails have become the current product cycle rather than an afterthought, with PubMatic shipping a five-component governance layer on August 5, 2026 that constrains what autonomous buying agents may do at the point of execution. Salesforce's answer is Agent Script plus per-agent permission inheritance, which places control at the level of business rules rather than at the level of spend.

The timing is not incidental. Dreamforce 2026 runs from September 15 to 17 at the Moscone Center in San Francisco under the theme of becoming an agentic enterprise, four days after this release. Salesforce has already announced Claudeforce with Anthropic, an arrangement disclosed on August 26, 2026 that put 37 prebuilt sales skills inside Claude for pilot customers and named Claude the default model across several Salesforce surfaces. The September 11 portfolio arrives as the applications layer of that architecture, and the pricing, packaging and edition structure attached to it were not part of the announcement.

Timeline

Summary

Who. Salesforce, with named customer deployments at Engine, Perk, Autism Queensland, Hibbett, Asana and Anthropic. Piper originated at Qualified, acquired in April 2026; Fin is Intercom, subject to a June 2026 acquisition agreement.

What. Seven prebuilt agents covering customer service (Casey), IT and HR service (Paige), shopping (Carter), outbound sales (Hunter), supply chain (Marshall), inbound pipeline generation (Piper) and customer experience workflows (Fin), plus a long-horizon runtime built on memory, durable execution and dynamic steering, and three platform additions: AI Skills, Multi-Agent Orchestration and Agent Optimizer.

When. Released September 11, 2026. Six agents generally available immediately; Hunter in pilot with general availability in November 2026; AI Skills and Agent Optimizer generally available in October 2026. Dreamforce follows on September 15 to 17, 2026.

Where. San Francisco, with agents operating across voice, SMS, WhatsApp, web chat, Slack, internal portals, websites and inboxes, connected to Customer 360.

Why. Salesforce is moving from selling an agent-building platform to selling configured agents for named jobs, backed by a volume metric it defines itself: 7 billion Agentic Work Units over two years and 3.2 billion in the second quarter, against 2.4 billion cumulative units disclosed in February 2026. The deployment statistics attached to each agent are vendor-supplied and carry no denominators, and the company's own benchmark research still records a sharp drop in agent reliability once tasks run past a single turn.