Dreamforce 2026 Agentforce recap: the work after launch

An agent going live is a milestone. The next customer asking something unexpected is where the useful work begins.

The Dreamforce 2026 Agentforce keynote focused on that longer journey. Salesforce organized the session around getting started, extending the experience, and improving it through observation. Fulton Bank, Southwest Airlines, and Vivint supplied the customer context; Agent Optimizer offered a preview of how the improvement process could become more connected. Watch the official Agentforce keynote

Our strongest takeaway is that an agent needs an operating rhythm. Someone must understand the work it is meant to do, see where the experience falls short, and turn those findings into a tested change.

Start with work a team can recognize

The Fulton Bank section demonstrated Agentforce Coworker helping a commercial leader examine opportunities and prepare coaching. It then showed a recurring review being turned into a reusable skill, with a specialized agent contributing to the result. Fulton Bank demonstration

That is a more useful starting point than asking a team to imagine every possible use of AI. A recurring review has recognizable inputs, a regular audience, and a result people can assess. The team can explain what belongs in it and what would make it misleading.

Turning that work into a reusable instruction set can also reveal hidden assumptions. One manager may prioritize a stalled relationship because a next step is overdue; another may be looking at a recent conversation. Before standardizing the review, those differences deserve a conversation.

We would choose a recurring task where people can compare the output with their own experience. That creates a practical basis for deciding whether to reuse, revise, or retire the approach.

Long-running agents need opportunities for people to steer

The Hunter demonstration followed a goal through a proposed plan, scheduled tasks, approval points, and drafted outreach. The session described a runtime that can retain context and adapt while work continues over time. Long-running agent demonstration

The important change is the unit of work. An answer happens in a moment. Re-engaging a relationship may involve several interactions, a delay, new information, and a different decision next week.

That makes visibility into the plan valuable. A seller should be able to understand what is expected to happen next and change direction when the relationship changes. A goal should not keep driving activity after its original reason has disappeared.

We would test that explicitly. Introduce a new piece of information midway through a trial and inspect the resulting plan. For example, the customer may ask to pause a discussion or name a different stakeholder. A useful agent should help the team respond to the change, rather than make the old plan run faster.

Hunter remains a pilot with general availability targeted for November 2026, according to Salesforce’s dated product announcement. The keynote demo does not change that release status. Agent availability

Southwest connected a helpful conversation to a defined process

Southwest’s section moved from an information assistant into an account-management example. The name-change demonstration included checks, document handling, a MuleSoft action, and a case for follow-through. The session also showed how separately maintained agents could contribute to a unified customer experience. Southwest demonstration

The lesson is in the sequence. A customer experiences one request, even when the work crosses several technical boundaries. Asking the customer to repeat information at each boundary makes the organization’s structure their problem.

A well-designed workflow preserves what has already been established while applying the next required step. That depends on clear interfaces between systems and clear responsibilities between teams. Our revenue systems architecture work uses the same principle: follow the customer’s need across the systems that must satisfy it.

Southwest’s Justin Bundick also emphasized a focused starting point and learning from real conversations. His account connected ongoing improvements with changes in how business and technical teams work together. Customer discussion

That is a staffing and ownership question as much as a product question. A queue of insights is useful only if a team can decide which ones matter and act on them.

Vivint showed why a missing capability can be useful evidence

The Vivint segment examined a pattern in customer interactions: people wanted a device reboot that the agent had not yet been configured to perform. The demonstration used that observation to introduce an action and test the change. Vivint observation and improvement example

That is a concrete improvement loop. A customer request reveals a gap. The team determines whether the agent should handle it. A change is made and tested. The next set of interactions tells the team whether the experience improved.

We would keep the original customer problem attached throughout that process. Otherwise, it is easy to declare success because a new action was deployed while losing sight of whether the customer’s task became easier.

The measure should reflect the outcome the customer wanted, including any follow-up work. A conversation ending is not necessarily a problem being resolved. That distinction belongs in the team’s definition of success before results appear on a dashboard.

Agent Optimizer joined the steps together

The Agent Optimizer preview began with an observed refund-handling issue. It carried that context into diagnosis, a proposed plan, configuration changes, tests, and a rollout decision. The presenter showed a person reviewing the work throughout the sequence. Agent Optimizer preview

Salesforce Ben’s practitioner recap also identified the optimization demonstration as a particularly practical part of the session. That corroborates the emphasis, while the official recording supplies the evidence for what was shown. Practitioner recap

The value of connecting these steps is traceability: the proposed change should remain attached to the problem it is intended to fix. A reviewer needs to understand why the change exists and what the test results establish.

There is still a difference between a convincing demonstration and an available production capability. Salesforce’s September 11 announcement targets Agent Optimizer general availability for October 2026. The keynote calls this a preview. Official availability

We would use the preview to shape an evaluation, while checking actual access and release details before including it in delivery commitments.

Give the improvement work a home

The practical next step is to bring the business owner and the people maintaining the agent into the same review. Start with a small set of actual interactions. Identify the recurring point of friction, decide what a better outcome would be, and select a change the team can evaluate.

That review should produce a decision, not simply more observations. Some issues need a new capability. Others need clearer source information or a better handoff to a person. The correct response depends on the work.

Our AI for business and revenue operations consulting work connects those decisions to practical outcomes. Launch gives a team something to learn from. A repeatable way to improve it is what makes the investment keep working.

About CRM Hacker

We help teams connect Salesforce, revenue operations, and AI around the way their business runs. Our Salesforce consulting work includes the workflows and operating practices needed after implementation.

If your agent roadmap needs a clearer path from demonstration to everyday value, Contact CRM Hacker. Let’s untangle this. The initial conversation is free; audits and implementation are scoped and paid.