Seven Agents with First Names and a Memory That Outlives the Chat: Inside Salesforce's Long-Horizon Agentforce

For two years Salesforce has sold Agentforce as a platform: here are the tools, here is the atlas reasoning engine, go build your own agents. On September 11, 2026, the company changed tack. It released a portfolio of seven prebuilt, “job-ready” AI agents — each with a first name and a job description — and, more consequentially, a new long-horizon runtime designed to let an agent hold a goal across days and weeks rather than a single conversation window.

Six of the seven agents are generally available immediately. The seventh, an outbound sales agent named Hunter, remains in pilot until November 2026 — and Hunter is the one running on the new runtime, which makes it the most interesting thing in the release.

Seven agents, six live now

The portfolio covers the full span of front-office and back-office work:

• Casey — customer service resolution across voice, SMS, WhatsApp and web chat, with prebuilt handling for FAQs, returns, account management and escalation to a human.
• Paige — IT and HR requests, reachable through Slack, internal portals and existing employee tools.
• Carter — the commerce agent: helps shoppers find and compare products and closes the transaction through in-chat checkout.
• Hunter — outbound sales; works a pipeline from research through outreach “over weeks and months.” Pilot now, GA in November.
• Marshall — supply chain and back-office orchestration, sold on “deterministic execution” and an audit record of every action taken.
• Piper — engages, qualifies and converts inbound leads on websites and inboxes into B2B pipeline.
• Fin — resolves complex customer-experience workflows across channels, running on a customer-operations agent called Operator and Fin Apex models custom-trained for CX work.

Every agent connects to Customer 360 and inherits existing customer records, business rules, permissions and the security model. Customers can rename any agent to match their own branding, and behaviour can be specified through Agent Script — the open-source language Salesforce published earlier this year that mixes model reasoning with deterministic rules, so certain decisions follow fixed logic rather than inference. That combination is the quiet headline: a system that reasons freely is hard to audit, and the back-office agent in particular is being sold on its audit record.

The runtime is the real story

The packaging gets the attention, but the infrastructure announcement matters more. Salesforce built a long-horizon runtime for Agentforce, and Hunter is the first agent to run on it. The company says more agents will migrate onto it over time, and that customers will eventually build long-horizon agents of their own.

Three capabilities sit underneath:

1. Memory carries context and progress between sessions, so a plan survives the end of an interaction.
2. Durable execution keeps that plan running and lets the agent resume or correct course when circumstances change.
3. 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 the 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.

Notably, the announcement does not specify 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 February 2026 working paper, separated a single AI agent that acts with some autonomy from coordinated multi-agent systems pursuing objectives over extended periods with minimal supervision. Sustained goal pursuit over weeks is the second category — the one where oversight, liability and specification questions are least settled.

Two of these agents arrived by acquisition

Piper and Fin did not originate inside Salesforce. Piper is the inbound sales product built by Qualified, founded in 2018 by former Salesforce executives Kraig Swensrud and Sean Whiteley. According to Salesforce’s quarterly SEC filing, the company acquired Qualified in April 2026 for approximately $1.2 billion, roughly $1.1 billion of it cash, recording $954 million of goodwill.

Fin is Intercom. The same filing records a June 2026 agreement 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 homegrown portfolio is a product decision — neither the announcement nor the filings describe how deeply either system has been rebuilt on Salesforce’s own stack.

The numbers, and what they leave out

Salesforce framed the release with volume rather than revenue: 7 billion Agentic Work Units delivered across Agentforce and Slack over two years, including 3.2 billion in Q2 alone. An Agentic Work Unit — the metric Salesforce introduced at its February 2026 results — is one discrete task accomplished by an agent: a prompt processed, a reasoning chain completed, a tool invoked. The company positions it explicitly against token counts, arguing tokens measure consumption rather than completed work.

The arithmetic is the story. Cumulative units moved from 2.4 billion in February to 7 billion roughly two quarters later, and a single quarter now accounts for 3.2 billion — more than four times the quarterly figure disclosed seven months earlier. But the metric remains defined and counted by the vendor, with no external audit and no published breakdown by agent type. A tool invocation that fails still counts as work performed.

The customer statistics carry the same caveat. 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 covers 90% of core shopper journeys after a six-week deployment. Anthropic resolves 79% of the conversations Fin sees without human involvement. All are vendor-supplied, and none carries a denominator, a measurement window or a definition of “resolution” — a rate that depends entirely on which contacts enter the funnel.

Measured against Salesforce’s own research

The most useful counterweight comes from Salesforce AI Research itself. Its CRMArena-Pro benchmark, published in June 2025, found leading language-model agents succeeded in 58% of single-turn business tasks but only 35% of multi-turn ones across 19 business tasks and 4,280 queries. Workflow execution was the most tractable skill, exceeding 83% single-turn, while confidentiality awareness was a consistent weakness across every model tested.

That study measured agents inside a conversation. The long-horizon runtime extends the horizon to days and weeks, multiplying the 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 mid-flight.

Context: Dreamforce in four days

The timing is not incidental. Dreamforce 2026 runs September 15–17 at the Moscone Center in San Francisco under the theme of “becoming an agentic enterprise” — four days after this release. It also follows the Claudeforce arrangement with Anthropic disclosed on August 26, which 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; pricing, packaging and edition structure were not part of the announcement.

Three platform additions accompany the portfolio: AI Skills inside Agentforce Coworker (teach a task once, reuse across the workforce; GA October), Multi-Agent Orchestration for routing work between specialised agents (GA now), and Agent Optimizer for building, testing and trace-analysis of agents (GA October).

For buyers, the pitch is simple: prebuilt agents shorten the distance between purchase and deployment. For the industry, the signal is sharper — the competitive frontier in enterprise AI has moved from “can agents reason” to “can agents be trusted to hold a goal for a month.” Salesforce is betting the answer is yes. Its own benchmark data suggests the second half of that bet is still unproven.