Salesforce Job-Ready Agents and Long-Horizon Work
Seven named agents got the headline, but the long-horizon runtime underneath is the part a small team can actually copy.
Salesforce Job-Ready Agents and Long-Horizon Work
On 11 September 2026 Salesforce announced seven named Agentforce agents and, alongside them, a long-horizon runtime that lets an agent pursue a goal across days or weeks instead of finishing inside a single session. The named agents are the marketing. The runtime is the part worth reading, because the three capabilities it is built from are ideas any small team can implement without buying anything. Details below come from the Salesforce announcement.
What the Salesforce job-ready agents actually do
Six of the seven agents are generally available now: Casey for customer service across voice, SMS, WhatsApp and web chat, Paige for IT and HR requests through Slack and portals, Carter for product discovery and checkout, Marshall for back-office and supply chain orchestration with audit records, Piper for inbound lead qualification, and Fin for multi-channel customer workflows. Hunter, the outbound sales agent, is in pilot with general availability planned for November 2026.
Hunter is the only one running on the new runtime today. Salesforce says more of the portfolio will move onto it over time, and that customers will be able to build long-horizon agents themselves.
The three ideas inside the long-horizon runtime
Salesforce describes the runtime as three capabilities. Stripped of the branding, they are a decent specification for any agent meant to survive longer than one conversation:
Capability | What it means | What it replaces |
|---|---|---|
Memory | Context carries across sessions rather than resetting each run | Re-pasting the same brief every morning |
Durable execution | A plan persists and survives interruption, with room to course-correct | A crashed run that starts from zero |
Dynamic steering | Behaviour adapts to feedback mid-flight, not just at the next prompt | Killing the run and rewriting the prompt |
None of that requires a Salesforce licence. Memory across sessions is a storage problem before it is a model problem, which we walk through in how AI agents remember between sessions. Durable execution is the same idea a job queue has used for twenty years, applied to a plan instead of a task. Dynamic steering is the hardest of the three to get right, because an agent that adapts to every piece of feedback drifts, and one that adapts to none is a batch script with a chat interface.
If the underlying concept is unfamiliar, what a long-horizon task in AI actually is is the background piece. The distinction that matters for your own build is whether you need an agent at all or just a scheduled workflow, which is the question in AI agent versus AI workflow.
About those customer numbers
The release carries a set of striking figures: 60 percent of one customer's sales pipeline built by Hunter, 70 percent of another's administrative requests resolved by Paige, 90 percent of a retailer's core shopper journeys handled by an agent, 79 percent autonomous resolution on Fin at Anthropic. Every one of those is a vendor-reported number about a vendor's own product, supplied by customers the vendor selected.
That does not make them false. It makes them unaudited, with no stated definition of what counts as resolved or as pipeline, and no denominator. Read them as the ceiling a well-matched deployment reached, not as what a typical rollout produces. The honest version of this figure for your own build is the one you measure after you ship.
Why this release matters beyond Salesforce
The interesting signal is not that a large vendor shipped named agents. It is that the hard problem has visibly moved. A year ago the open question in agent products was whether the model could complete a task at all. The framing now is that single-task completion is assumed and the difficulty is continuity: holding a goal across an interruption, across a week, across a change of instruction halfway through.
That shift is what makes this worth a small team's attention even if enterprise CRM is nowhere near their world. The primitives are the same at any size, and the failure modes will be too: an agent with memory and no steering quietly compounds a bad assumption for a week. If you are wiring several of these together, multi-agent orchestration covers the coordination layer that sits above all of this, and our running guide to following AI releases is where this one sits in the wider run of the week.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
