On September 11, Salesforce announced a new portfolio of "job-ready" AI agents and, buried underneath the product names, something more structurally important: a long-horizon runtime that lets an agent pursue a goal across days and weeks instead of finishing when the conversation ends. Salesforce's example is a seller telling an agent named Hunter to "rescue at-risk deals before the end of the quarter." Hunter turns that into a plan, works it over the following weeks, adjusts as new information comes in, and checks back in with the human when it hits a decision point that needs approval.
That is a different kind of software than the AI most small businesses have used so far. A chatbot answers a question and forgets you the moment the window closes. Even most "AI agents" today complete one task — draft this email, summarize this call, update this record — and stop. A long-horizon agent keeps a goal alive across sessions, days, and weeks, the same way a human employee would keep working a project without needing to be re-briefed every morning. Whether or not you run Salesforce, this is the direction every AI vendor is now headed, and it changes what "supervising an AI agent" actually requires.
What Salesforce Actually Launched
Salesforce's announcement bundles two distinct things. The first is a set of prebuilt, named agents — Casey (customer service), Paige (IT/HR), Carter (shopper), Hunter (outbound sales), Marshall (supply chain and back office), Piper (inbound pipeline), and Fin (customer experience) — that ship with the skills, actions, and data models for a specific job already built in, connected to a company's existing Salesforce data. Most of these are generally available now; Hunter is in pilot with general availability planned for November 2026.
The second, more consequential piece is the long-horizon runtime itself, which Salesforce says rests on three capabilities: memory that preserves context and progress across sessions so work doesn't reset between conversations, durable execution that keeps a plan running and lets the agent resume or course-correct as circumstances change, and dynamic steering that adapts the agent's behavior based on a person's ongoing feedback. Salesforce reports real usage volume behind this shift — 7 billion "Agentic Work Units" delivered across Agentforce and Slack over two years, with 3.2 billion in the second quarter of 2026 alone — plus specific customer results: 60% of Perk's sales pipeline built by Hunter, 70% of Autism Queensland's admin requests resolved by Paige, and 90% of Hibbett's shopper journeys handled by its Carter deployment.
Why "Long-Horizon" Is a Different Category
We've written before about the shift from AI answering questions to AI taking real actions inside your business. Long-horizon runtimes are the next step past that: the difference between an agent that acts once and an agent that keeps acting, unsupervised, over an extended stretch of time. That distinction matters because the risk profile of a task doesn't just depend on what the agent can do in a single step — it depends on how many decisions accumulate before a human looks at the work again.
A draft-only agent that writes one email is easy to review before it sends. An agent working a deal for three weeks, sending multiple outreach messages, updating multiple records, and adjusting its own plan based on what it learns along the way is a fundamentally harder thing to supervise after the fact — which is exactly why Salesforce built guardrails directly into the runtime rather than leaving it to each customer to bolt on. Not every AI vendor will be that disciplined by default, so the responsibility for setting checkpoints often falls on you.
The Job-Ready Agent Roster
The specific named agents are worth knowing even if you don't buy Salesforce, because they preview where every CRM, help desk, and ERP vendor is headed:
- Casey resolves customer service issues across voice, SMS, WhatsApp, and web chat, with built-in handling for FAQs, returns, account management, and escalation to a human.
- Hunter researches accounts, builds outreach, and works a sales pipeline over weeks — the first agent running on the new long-horizon runtime.
- Paige resolves IT and HR requests across Slack, portals, and the tools employees already use.
- Marshall orchestrates back-office and supply chain processes with deterministic execution and an audit record of every action it takes.
- Piper works inbound leads across a company's website and inbox to qualify and convert them into pipeline.
The pattern across all of them: pick a job, not a feature. Small businesses evaluating any AI vendor's agent lineup — Salesforce or otherwise — should ask the same question these named agents answer for Salesforce customers: which specific job is this agent built for, what does it need access to in order to do that job, and what happens when it hits something outside its lane?
Thinking about piloting a long-running AI agent in your business?
We help small businesses pick the right first job for an agent, set checkpoints for multi-day work, and build the approval gates that keep long-horizon automation safe.
Book a Free Strategy Call →Why This Matters If You Don't Run Salesforce
Salesforce isn't alone in building agents this way, and it isn't even the only CRM-AI story we've covered this quarter — Salesforce's own partnership with Anthropic, Claudeforce, put Claude's reasoning directly inside the platform just weeks before this launch. OpenAI's GPT-6 Astra release the same month pushed hard on faster, more reliable computer-use — a different angle on the same underlying trend: AI models are shifting from single-turn helpfulness toward sustained, semi-autonomous work.
If your business runs HubSpot, a help desk tool, an accounting platform, or a project management system, expect a version of "long-horizon agent" to show up there within the next year, whether or not it's branded that way. The practical takeaway isn't "adopt Salesforce's agents." It's "assume every SaaS tool you already pay for is about to ask you to trust it with unsupervised, multi-day autonomy — and decide now what your answer to that request will be."
The New Governance Gap: Approvals Over Weeks
Most small businesses that have thought about AI governance at all have thought about it in terms of a single interaction: does this one email need review before it sends, does this one action need approval before it executes. Long-horizon agents break that model, because the risk isn't in any single step — it's in the accumulation of dozens of small decisions an agent makes on its own over a multi-week stretch before anyone checks in.
Salesforce's answer is to build "guardrails that define when it can act autonomously and when seller approval is required" directly into the runtime, which is the right instinct. Whatever platform you use, ask three questions before turning on any multi-day agent: What decision points automatically pause for a human? Is there a running log of every action the agent has taken, not just its final output? And can you interrupt the agent mid-plan if priorities change, without losing its accumulated context? These map directly onto the cost and permission controls we've recommended for any AI agent — long-horizon work just raises the stakes on getting them right, since a runaway agent now has weeks to compound a mistake instead of minutes.
2-Week Action Plan
You don't need to buy a long-horizon agent to get ahead of this. Do this instead:
- Days 1–3: Audit what "multi-session" AI you already have. Check whether your CRM, help desk, or scheduling tool has quietly enabled any agent feature that persists context across sessions. Many roll out as opt-in defaults you never explicitly reviewed.
- Days 4–7: Pick one candidate job, not one candidate tool. Following our pilot playbook, identify a single recurring, multi-step job — following up on stalled deals, working an admin queue, chasing overdue invoices — that would benefit from an agent that keeps working after the conversation ends.
- Days 8–10: Define your checkpoints before you turn anything on. Write down, in advance, which actions in that job require human sign-off and which are safe to run autonomously. Don't let the vendor's defaults make that decision for you.
- Days 11–14: Ask every vendor pitching an "agent" the same three questions. What pauses for approval, is there an action log, and can you interrupt it mid-plan. If the answer is vague, that's your answer about whether to turn it on yet.
Bottom Line
Salesforce's launch matters less because of the specific named agents and more because it marks AI's shift from single-turn helper to standing employee — something that pursues a goal across weeks, adjusts its own plan, and checks in only when it needs to. That's genuinely useful for the right jobs. It also means the old governance question — "should I review this one output before it goes out?" — isn't sufficient anymore. The new question is "how much can accumulate before I look again?" Get that answer right before you turn any long-horizon agent loose on real customer relationships or real money.
If you want help figuring out which job in your business is ready for a long-running agent, or building the checkpoints to supervise one safely, book a free strategy call at apolloagent.ai. We'll help you separate the genuinely useful automation from the stuff that just sounds impressive in a press release.