If your Salesforce renewal deck still says “AI chatbot for service,” throw the slide out before Dreamforce opens. On September 11, 2026, Salesforce did not ship another answer engine. It shipped seven named job packs and a runtime that keeps working after the chat window closes. That is an org chart drop disguised as a product launch.
The buy for marketing, growth, brand, and ops is no longer “can the bot answer.” It is which role you are handing a multi-week goal to, whose voice it speaks in, and who owns the SLA when it keeps going without you.
What actually changed
Salesforce’s newsroom post (September 11, 2026), mirrored the same day on the UK newsroom, named seven “job-ready” Agentforce agents tied to Customer 360:
- Casey: help agent for service across voice, SMS, WhatsApp, and web chat. GA now.
- Paige: IT and HR employee service across Slack, portals, and existing tools. GA now.
- Carter: shopper agent for discovery, comparison, Q&A, and in-chat checkout. GA now.
- Hunter: outbound sales from research to outreach over weeks and months. Pilot now. GA November 2026.
- Marshall: supply chain and back-office orchestration with deterministic execution and an audit record. GA now.
- Piper: inbound pipe gen across websites and inboxes for B2B sales and marketing. GA now.
- Fin: CX workflows across channels, powered by Operator and Fin Apex models. GA now.
AI Weekly (Alexis Dufresne, September 12) and Unite.AI (Aiden Cross, September 11) independently confirm the same roster, the six-now / Hunter-later GA split, and the packaging move: start from a job, not a blank canvas. Enterprise DNA (September 11) and PPC Land (September 14) put the timing in front of Dreamforce 2026 (September 15 to 17, San Francisco). This is the applications layer walking into the keynote week, not a quiet docs update.
Under the names sits a scale claim Salesforce wants every renewal deck to absorb. Across Agentforce and Slack, the company says it has delivered 7 billion Agentic Work Units, including 3.2 billion in Q2 alone. Unite.AI notes Salesforce introduced the AWU metric in February 2026 as one discrete task accomplished by an AI agent. AI Weekly and PPC Land both stress the same caveat you should keep in the room: the metric is vendor-defined and vendor-counted. Treat it as platform activity signal, not audited customer outcome.
The second half of the announcement matters more for GTM than the names alone. Salesforce built a long-horizon runtime so agents can pursue goals across days and weeks instead of finishing one interaction. Memory keeps context across sessions. Durable execution keeps plans running and lets the agent resume or course-correct. Dynamic steering adapts behavior from user feedback. Hunter is first on that runtime. More of the portfolio follows. Customers will eventually build their own long-horizon agents on Agentforce. That trio (memory, durable execution, dynamic steering) is corroborated across Salesforce’s post, AI Weekly, Unite.AI, and PPC Land.
Around the agents sit platform pieces with their own dates. Multi-Agent Orchestration is GA now and routes work across specialized agents when a job crosses roles or journey stages. AI Skills in Agentforce Coworker is pilot now, GA October 2026. Agent Optimizer is GA October 2026. Agent Script remains the open-source control language that mixes reasoning with deterministic rules. Customers can rename agents so they look like brand extensions, not Salesforce mascots.
Salesforce also published customer proof points: Engine’s help agent resolving 50% of chat inquiries, Perk attributing 60% of pipeline to Hunter, Autism Queensland at 70% admin resolution via Paige, Hibbett AI on 90% of core shopper journeys in six weeks, Asana’s Piper at 4x conversation volume with ~45-day average deploy, Anthropic at 79% autonomous resolution on conversations Fin sees. Label those as vendor-supplied. AI Weekly and PPC Land both say the same: no third-party audit, thin denominators, activity mixed with outcome. Useful as adoption theater. Useless as your ROI model until you rebuild the math on your own funnel.
The mechanism: the job pack is the control surface
Strip the Dreamforce gloss and the product boundary is simple.
A chatbot SKU sells a capability: answer, deflect, draft. A job pack sells a role: outbound seller, inbound pipe gen, shopper, CX closer, back-office orchestrator. The pack arrives with skills, actions, and data models for that job. It inherits Customer 360 context. It can be renamed into your brand. It can be steered by Agent Script. Once Multi-Agent Orchestration is live, work can hop across those roles as one coordinated sequence.
That is why the long-horizon runtime is not a feature footnote. Hunter’s worked example is a seller asking it to rescue at-risk deals before quarter end. The agent turns that into a measurable goal, builds a plan, chooses tools and context, and applies guardrails for when it acts alone versus when it needs approval. Memory means the plan survives the meeting ending. Durable execution means the plan is still running on Tuesday. Dynamic steering means your AE can yank the wheel without rebuilding the agent. Weeks of outreach are not “a longer chat.” They are a goal with owners, approvals, and brand exposure while you sleep.
For marketing and revenue teams, two of the seven land directly on the path: Piper on inbound websites and inboxes, Carter on commerce discovery and in-chat checkout. PPC Land’s September 14 package is blunt that pipeline generation, chat commerce, and service deflection now arrive as configured products rather than components to assemble. The operator question flips from “what do we build” to “what role are we staffing with a persistent goal.”
What operators get wrong
Mistake one: buying this like another chatbot SKU. If procurement, enablement, and creative still treat Casey or Piper as “the AI widget on the site,” you will under-scope brand voice, escalation rules, and who gets paged when the agent keeps working for three weeks. Named agents are headcount-shaped. Budget and ownership should look like that.
Mistake two: assigning no human owner because “it is autonomous.” Long-horizon work without a named human sponsor is how brand voice drifts, offers get overpromised, and SLAs become nobody’s problem. Hunter can pursue a quarter-end rescue plan. Someone still owns the deals, the sequence tone, and the approval thresholds.
Mistake three: letting service, sales, and marketing each “try an agent” without a shared routing map. Multi-Agent Orchestration is GA. That means a lead can bounce from Piper to Hunter to Casey without your org chart noticing. If those teams do not share definitions for qualified, escalated, and brand-safe, orchestration will amplify inconsistency instead of fixing handoffs.
Mistake four: confusing vendor proof with your funnel math. Perk’s 60% pipeline figure and Asana’s 4x conversation volume are Salesforce release stats. PPC Land flags that pipeline built is not pipeline closed, and conversation volume is activity. Rebuild the scorecard before you put Hunter or Piper into the board pack.
Mistake five: ignoring GA timing as an operating constraint. Six agents are buyable now. Hunter is pilot until November 2026. AI Skills and Agent Optimizer land in October. If your Q4 plan assumes long-horizon outbound at full production grade today, you are planning fiction. Pilot Hunter with clear guardrails. Put Piper and Carter on the nearer calendar if inbound and commerce are the pain.
Mistake six: treating rename-the-agent as branding done. Salesforce says customers can give agents their own names. That does not write your tone guide, disclosure copy, or escalation script. Brand and legal still have to ship the voice pack before the agent talks to a prospect for fourteen days.
Mistake seven: skipping the Casey versus Fin question. AI Weekly asks how Salesforce’s own CS agent and the Fin line divide the floor. If you run both narratives into one CX org without a RACI, you will double-pay for deflection and still not know who owns the hard tickets.
Second-order effects
GTM orgs will start arguing about agent headcount the way they argue about SDR headcount. Enterprise DNA’s read is directionally right even when it sells its own platform: naming the role makes ROI and change management concrete. “We are staffing Hunter on outbound while sellers close” is a different conversation than “we are implementing AI.” Expect budget lines labeled by agent name by next planning cycle.
Agencies and RevOps will inherit a new deliverable: the agent RACI. Who owns Piper’s qualification rules. Who owns Carter’s checkout promises. Who owns Hunter’s email voice. Who owns Marshall’s audit trail when finance asks what changed. The agencies that only sell prompt libraries will lose to the ones that sell role design, voice packs, and escalation matrices.
Measurement will fight AWUs. Salesforce wants work units as the platform story. Marketers still need pipeline, conversion, CSAT, and cost-to-serve. Until someone maps AWUs to those outcomes in your instance, do not let the 7B headline substitute for a dashboard you trust. AI Weekly’s open question on how AWUs meter against existing per-conversation Agentforce pricing is still open in public coverage. Do not invent an answer in the CFO deck.
Governance pressure rises because weeks of autonomous pursuit multiply state transitions. Salesforce’s own earlier CRMArena-Pro research (cited in PPC Land’s package) showed multi-turn business tasks succeeding far less often than single-turn ones. Long-horizon runtime extends the horizon. It does not erase that reliability cliff. Treat Optimizer (October GA) and session-trace review as mandatory ops, not nice-to-have.
Dreamforce week will amplify the packaging. Expect demos of named agents as coworkers with jobs. Your job is to walk out with a role map, not a screenshot of a cheerful agent avatar.
What to do this week
- Draw the agent org chart before you draw the architecture diagram. One page. Seven rows if you are on Salesforce, or the subset you will actually touch. Columns: job, human owner, brand voice owner, SLA, GA status, systems it can touch. Put Piper and Carter on the marketing side of the page even if IT owns the tenant.
- Pick one multi-week goal, not seven pilots. If you have Hunter pilot access, define one quarter-end style objective with approval thresholds in writing. If you do not, pick Piper or Carter with a single journey (inbound qualify, or one commerce path) and a kill switch. Duration beats breadth.
- Rewrite brand and legal for persistent agents. Disclosure, tone, offer boundaries, escalation copy, and what the agent must never promise. A fourteen-day Hunter sequence is a campaign, not a widget. Treat it like one.
- Force a shared definition meeting across sales, marketing, service, and RevOps. Qualified lead. Brand-safe reply. When Multi-Agent Orchestration may hand off. Who gets the transcript. Ninety minutes. Output is a one-pager, not a Slack vibe.
- Build a scorecard that is not AWUs. For Piper: qualified pipeline, speed-to-lead, meeting rate. For Carter: assisted conversion, refund and complaint rate on agent-closed carts. For Casey or Fin: containment, reopen rate, CSAT on escalations. Map any Salesforce work-unit story to those lines later.
- Calendar the GA reality. Hunter November. Skills and Optimizer October. Dreamforce sessions this week for demos, not for pretending production is already solved. Update the Q4 roadmap so long-horizon outbound is labeled pilot until GA evidence exists in your org.
- Ask pricing and metering questions out loud. How AWUs relate to your current Agentforce conversation SKUs, whether long-horizon carries a premium, and how Casey and Fin are licensed if both appear in your CX stack. AI Weekly flagged those as open. Get answers from your AE before you forecast savings.
Sharp close
Salesforce just turned Agentforce from “build an agent” into “hire a named job.” Six of those jobs are for sale now. One is learning to chase goals across weeks. Operators who still shop this like a chatbot SKU will mis-assign ownership, brand voice, and SLA while the agent keeps working.
Named agents are org charts. Staff them like you mean it, or do not turn them on.
