Agentforce: Sales Agents That Work Toward Goals Across Weeks
Agentforce long-running agents use memory and persistent execution for sales goals across days or weeks. Pilot reversible follow-up work with clear owners; verify recovery, approvals, costs and exceptions rather than assuming better sales results.
What's changed: Goals can continue to progress after the conversation ends
Salesforce announced the Agentforce long-term runtime on September 14, 2026. The change is that agents can continue working on goals across days and weeks, rather than just completing the current question and answer.
For sales, customer responses, demand changes, and internal decisions often don’t come together at the same time. A single-turn response answers the immediate question, while an ongoing task requires remembering what has been done, what is still waiting for, and who needs to make the next decision.
Hunter is the first agent to adopt this runtime. Citing the example of rescuing a high-risk business opportunity, officials described how it turns goals into plans and adjusts actions as new information emerges, while retaining control of salespeople.
What problems do memory, persistent execution, and dynamic guidance each solve?
Long-term work requires connecting context, execution progress, and new instructions from people. Memory answers "what happened before", persistent execution takes over "where the work is going", and dynamic guidance allows salespeople to change subsequent directions.
For example, if a client has raised new concerns, the agency should move forward based on the information it is allowed to use rather than sticking to an outdated plan. There is still a difference between being able to save information and being able to use it correctly. During the pilot, it is necessary to check whether the updated actions meet the current requirements.
| Ability | The role of the official description | What should enterprises observe when conducting pilot projects? |
|---|---|---|
| Memory | Preserve context and progress across sessions | After ending the conversation, continue the task and check whether the customer's new requirements are consistent with the existing progress. |
| Durable execution | Allow the plan to continue over time and recover or correct as circumstances change | After interruption, check whether completed actions have been redone and whether unfinished items can still be tracked. |
| Dynamic steering | Adjust agent behavior based on user feedback and instructions | After changing goals or discontinuing an arrangement, watch to see if follow-up actions take a turn, and leave a verifiable record. |
Which sales tasks are suitable to be handed over to agents for continued follow-up?
For B2B service companies that already use Salesforce and have a long sales cycle and many touch points, they can first consider internal information collection and follow-up suggestions: continue to collect the changes in business opportunities that are allowed to be read, and hand over the matters that need sales processing to the hands of the responsible person. Such tasks make it easier to state goals and withdraw recommendations.
If customer data is scattered, records conflict with each other, or there is no clear person responsible for the business opportunity, the data and responsibility should be resolved first. Otherwise, the task will run longer and incorrect judgments may be carried into more subsequent actions.
Stalled or high-risk opportunities can be targeted for piloting, but “continue follow-up” should not automatically equate to “continue outreach.” After the customer explicitly refuses contact, the needs have changed, or the sales is taken over, the corresponding stop or approval branch should be transferred instead of increasing contacts to reflect the progress.
| current conditions | Tendency to pilot scope | What information would change the arrangement? |
|---|---|---|
| Information is available, goals are clear, and responsible persons are clear | Let agents sort through changes and recommend next steps, starting with rollback items. | The quality of suggestions and exception records that have been manually checked can determine whether to add authorization actions. |
| Reading of data is allowed, but no approval rules have been defined for external communication. | Keep drafts and to-dos, and then contact customers after sales confirm them. | Evaluate outreach pilots only after the recipients, content, sending permissions, and stopping conditions are clear. |
| Missing records or conflicting opinions | The person in charge will first complete the basis for judgment, and the continued implementation will be suspended. | Once the sources of information and current customer intent are clarified, re-target the mission. |
Permissions and manual takeover should be determined by specific actions
While the task continues to run, customer information may change and the authorizer may change his or her mind. Enterprises should implement "can do it by themselves" and "requires sales confirmation" into actionable actions, rather than giving agents a broad sales target and then letting it go.
The following are deployment design recommendations and are not a unified approval list prescribed by Salesforce for all enterprises. The closer we get to customer commitments, expense expenditures, or opportunity status changes, the more clearly authorized responsibilities and human intervention methods need to be.
- Customer data: Limit readable data according to task purpose; when encountering records that you are not authorized to access, hand them over to the person in charge for processing, and do not bypass permissions through other channels.
- Outreach actions: Explain the difference between drafting and sending, clarify which recipients and content require approval; and be able to pause sending when contact intentions change.
- Business opportunity stage: Check the basis required for stage adjustments, separate proposed updates and formal modifications, and avoid writing speculation into business progress.
- Budget and frequency: Agree on the usage boundaries of tool calls and reach, stop expansion actions when approaching the limit, and report remaining matters to sales.
- Target change: Keep a record of the new instruction taking effect, and check whether there are still pending actions when canceling the old arrangement.
- Stop and take over: After sales intervenes, you can clearly see what has been done, what has not been done, and what needs to be approved; the audit records should be sufficient to explain the basis for the action.
How does a pilot determine whether continued implementation is worth expanding?
First select follow-up tasks with descriptive goals and verifiable historical records, and use similar business opportunity status, customer contact conditions and judgment criteria to compare the existing methods with the performance after agent participation. Don’t allocate more active customers to agents and explain the difference as an automation effect.
Reliability requires observing changes as they occur: task interruptions, additional customer information, sales revisions, and delays in approvals can expose gaps that cannot be seen in a single round of Q&A. When incorrect outreach or repeated execution occurs, first tighten permissions and locate status problems, and then consider restoring the pilot.
Missed follow-up, amount of manual review, incorrect actions and quality of opportunity progression should be looked at together. If reducing omissions is accompanied by a large number of error corrections, it may not be worth expanding; after the quality is stable, the complete cost including review, exception handling and maintenance can be calculated. The transaction effect still needs to be verified with the company's own records and cannot be directly inferred from the continuous operation capability.
Separate Capability Claims from Purchasing Decisions
The confirmed long-term capability range is subject to official announcements and Hunter’s launch. The access of other agents and the long-term agents built by enterprises are subsequent expansions, and it cannot be concluded that all Agentforce agents now have the same capabilities.
Being able to continue the plan does not mean that sales can be completed unattended. Before going online, the actual available scope, access conditions, and permission settings of the proposed functions should be checked.
Prepare existing follow-up processes, available customer records, observations of stalled opportunities, a list of authorized actions, and persons responsible for taking over. Combining these inputs can determine which types of tasks should be piloted first, which aspects should be left to human decision-making, and when situations arise that should be reduced in scope or stopped.
Common Problems
If a customer doesn’t respond for a long time, should the agent keep following up?
Silence alone is not enough to support sustained outreach. It should be handled based on the existing contact intention, sales judgment and task stop rules; when it is impossible to judge, return the "waiting for reply" to the person in charge for decision. The pilot can verify that the agent correctly retains the wait state without converting the wait into a duplicate send.
Sales has taken over the opportunity, can it be handed back to the agent later?
You can use the rehandover as a pilot design, but first explain what changes occurred during the sales intervention and what the agent will be responsible for next. Missions should be resumed with the latest approved records and clear new goals to avoid using the old plan before the takeover; specific handover capabilities need to be checked in the proposed configuration.
Does a controllable budget mean that long-term agency is worth investing in?
Budgets only constrain expenditures and cannot account for output. You should also judge whether the recommendations help sales make effective decisions and whether the manual processing burden is acceptable. If the saved follow-up time is offset by error correction, you can first reduce the types of tasks instead of just relaxing the budget and continue running.