Agentforce Can Work for Weeks—but Hunter Is the First Long-Horizon Agent
Salesforce detailed its long-horizon Agentforce runtime on September 16, 2026, describing a system that can keep pursuing the same goal across sessions, days or weeks. The September 16 runtime explainer identifies Hunter, an outbound sales agent, as its first implementation and shows human approval remaining in the path when execution departs from an agreed plan.
Availability is narrower than the platform vision. Salesforce’s September 14 portfolio announcement lists Hunter as a pilot, targets general availability for November 2026 and presents adoption by more agents—and eventually customer-built long-horizon agents—as future expansion.
The runtime preserves a goal, not just a conversation
Memory carries context and progress from one session to the next. In Hunter’s sales workflow, that state can include the objective, completed work, new engagement signals and upcoming tasks. Ending a conversation therefore does not reset the assignment.
Durable execution keeps the plan active over time and allows work to resume or change course when circumstances shift. Dynamic steering lets a user revise the agent’s guidance conversationally while execution is underway. Together, these capabilities make the continuing goal and its evolving plan—not an individual prompt—the unit of work.
The duration claim has important limits. The published example describes plans that can span days, weeks or even months, but the available material does not specify a maximum runtime, a recovery-time commitment or completion rates for multiweek goals. It defines an intended operating model rather than demonstrating that every extended assignment will finish successfully.
An at-risk deal shows what the runtime retains
The detailed scenario begins with a seller asking Hunter to re-engage at-risk deals before quarter-end. Hunter examines open opportunities, incorporates engagement signals from Data 360 and Slack, and converts the request into a measurable, time-bound objective. It then prepares tasks, timing and guardrails that determine when it may proceed and when it must return to the seller.
The memory layer retains that objective and accumulated progress across later sessions. Durable execution keeps scheduled work moving and provides a point from which the plan can resume. If the seller changes the preferred tone for a category of recipient, dynamic steering can update the guidance without discarding the broader objective.
The scenario’s clearest escalation point follows a failed outreach step. After an email bounces because the recipient is away, Hunter identifies another contact and drafts a replacement message. Sending to a different person falls outside the approved plan, so the agent pauses and asks the seller before continuing; it can also suggest a rule for handling similar responses later.
This is a vendor-supplied illustration, not a published reliability test. Even so, it defines the claimed division of labor: Hunter can detect a failure, prepare an alternative and retain the surrounding goal, while a consequential change remains subject to human review.
Human control remains in the execution path
The seller sets the initial objective, reviews the generated plan and can revise it before activation. The plan contains guardrails, and Hunter requests access to the user’s email account before sending messages on that person’s behalf. Once execution begins, new guidance can alter the plan without erasing its history.
Approval gates also survive activation. In the bounced-email example, finding a possible replacement contact is part of recovery, but contacting that person requires renewed approval because the recipient has changed. Persistence extends the workflow’s duration; it does not grant the agent unrestricted authority.
The wider portfolio is designed to operate within customer business rules, permissions and security, while Agent Script can combine model reasoning with deterministic instructions. The available material does not provide a complete list of Hunter actions that always require approval or establish whether administrators can configure every escalation condition.
Current availability versus planned expansion
- Available as a pilot: Hunter is the first agent running on the long-horizon runtime.
- Scheduled milestone: general availability for Hunter is targeted for November 2026, making it a forward-looking release date rather than a current capability.
- Planned expansion: additional agents are expected to adopt the runtime, but no product-by-product timetable has been published.
- Future platform capability: customers are expected to gain the ability to build and customize their own long-horizon agents, with no public delivery date yet.
SiliconANGLE’s launch coverage also describes the runtime as initially limited to Hunter, with integration into more agents planned later. That distinction narrows the announcement’s practical scope: the architecture is positioned for Agentforce more broadly, but its first deployment is one outbound-sales agent rather than a generally available runtime for every agent.
The unanswered questions are operational
No operational results from long-running Hunter deployments appear in the cited material. Enterprise buyers therefore lack public completion rates for multiweek plans, details about resource consumption over time, documented behavior after prolonged interruptions and a comprehensive account of configurable approval policies.
The confirmed development is a runtime designed to preserve goals and revise plans across extended work, with Hunter as its first user and human review retained before activation and at material plan changes. Hunter’s planned general release is the next stated milestone; wider agent support and customer-built long-horizon workflows remain future capabilities without published delivery dates.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.