Dreamforce covered more ground this year than any one person could realistically sit through. I made it to a number of sessions in person, then caught up on the rest online afterwards. This is my attempt to pull together the parts that matter for teams trying to figure out what changes after the announcements are over. Rather than list every feature Salesforce launched, I have focused on the things that could change how people work day to day.
A few themes kept coming up across almost every product.
One is that Salesforce increasingly does not expect you to spend all day inside Salesforce. AIforce is the broader idea behind that, with Claudeforce inside Claude, Slackforce inside Slack and Agentforce Coworker inside Lightning. The idea is fairly simple: people work where they are most comfortable, while Salesforce data, context and permissions follow them.
MCP was everywhere as well. The Model Context Protocol has become the common way for AI assistants to connect directly to systems, and nearly every product announcement seemed to include an MCP server, a collection of skills, or both. The useful question now is what an agent can do across several applications at once.
The other recurring theme was control. Salesforce was fairly consistent on this point: let the model reason, but let your systems, permissions and business rules make the final call.
With that in mind, here is what stood out to me. If you are only here for one product, jump ahead:
Agentforce
Agentforce Coworker got most of the attention, but Skills may end up being the more useful feature in practice. If you have a process that somebody on your team runs every week, you can turn that process into a reusable skill that Coworker or another agent can run again later. That sounds fairly basic, but it addresses a problem almost every organization has: a surprising amount of important work still depends on somebody remembering exactly how they did it last time.
Agents can also maintain a goal for much longer than a single interaction. Instead of starting from scratch every time, they can keep track of where work stopped and continue as circumstances change. That is much more useful for something like pursuing a stalled opportunity over several weeks than it is for the usual one-question, one-answer AI demo.
Data 360
Data 360 is becoming much more accessible outside its own interface. Data mapping, identity resolution and segmentation can now be handled through the Data 360 MCP server from Claude, Slack or another MCP client, while still respecting the underlying Salesforce permissions. A conversational segment builder is also expected in the new year.
The Agent Context Engine was one of the more interesting announcements for me. We spent years trying to get customer data out of separate silos, then started creating another set of silos by giving every agent its own memory and context. The Context Engine is intended to make that information usable across agents, including agents outside Agentforce, without sending every piece of available context into every interaction. Instead, the system can provide the information that is relevant to the question being asked.
Lineage is another feature that may not get the same attention but is important. If an agent creates a segment using a field you were not expecting, you can trace where that field came from and where else it is being used. If agents are going to work with customer data, that is the kind of visibility I would want in place before giving them more autonomy.
Marketing Cloud
Palmata was one of the more unusual announcements. It looks at how AI answer engines describe your brand and helps identify the content influencing those answers. In the demo, two support pages were contributing to a more negative description of the company in AI-generated responses. That is the sort of problem marketing teams have not really had visibility into before. There are free reports available, and I would be tempted to run one before the next planning cycle simply to see what turns up.
The Campaign Agent is probably the bigger operational change. Instead of manually designing every step of a journey, you give the agent a goal and a budget. It can then determine the message, channel and timing for each person. Adaptive audiences also allow someone to move into another campaign, or receive nothing at all, depending on what the system thinks is appropriate. For marketing teams that have spent years building increasingly complicated journeys, that is a significant change in mindset. You are managing the objective rather than every branch in the journey.
On the B2B side, Piper handles inbound activity while Hunter works accounts in the background, with context shared between them. Piper can also present directly on a website using an existing slide library, shorten the presentation when somebody asks, and send the deck afterwards in the visitor’s language.
Sales Cloud
Momentum was probably the Sales Cloud announcement that stood out most to me. It listens to calls, emails and Slack activity and uses that information to update CRM records, including custom objects, qualification fields and deal reviews. Sales teams have been asking reps to keep all of that information updated manually for roughly twenty years, with mixed results. Removing that administrative work is potentially more valuable than a lot of the more impressive-looking AI demos.
Hunter is the outbound agent. It includes access to ZoomInfo, Demandbase and Apollo and can use previous closed opportunities to work out what your successful customers have in common. One part I liked was that it can show the evidence behind its conclusions, including quotes from previous deals, rather than simply presenting an ICP and asking you to trust it.
Claudeforce also ships with 37 sales skills and connects through MCP. One of the demos showed a revenue leader creating a sales command centre largely from a prompt. A few years ago, something similar would probably have meant a project involving a solutions consultant, several meetings and a fair amount of configuration.
Commerce
Carter is the new shopper agent and runs on Salesforce’s updated commerce search engine. Salesforce says the new search converts 13% better than the keyword search it replaces. What caught my attention, though, was that the search technology can also be deployed on Shopify, Adobe, Magento or a custom commerce stack. You do not have to move everything to Commerce Cloud to use it.
Commerce Connect takes the same idea in the other direction by making product catalogues available through third-party AI channels. ChatGPT and Gemini are live, with TikTok expected to follow in January. If more shopping journeys begin inside AI assistants rather than on a retailer’s own site, getting accurate product information into those channels becomes increasingly important.
Storefront Next adds industry templates and a developer toolkit that works with Claude or Cursor. Salesforce highlighted one brand that went live in 28 days. Point of sale is also expanding beyond iOS to Android and Windows, which removes one of the more practical hardware limitations for companies considering a move.
Slack
Slack CRM brings CRM functionality directly into Slack and is included with Business+. Slackbot is also becoming an MCP client, with connections into Sales Cloud, Service Cloud, Agentforce, Data 360 and Tableau. That is a bigger deal than it might sound. If sellers can get the Salesforce information they need from Slack, there is less reason for them to keep jumping between applications to check or update a record.
Slackforce Surfaces takes that further. You can describe the view you want and create a live, shared interface connected to your systems and governed by existing permissions. A deal-health view sitting in the channel where the team already works is probably more likely to be used than another document or dashboard somebody has to remember to open.
Slack Code was also interesting. Coding agents can work inside a channel while non-developers watch what is happening, read the changes and redirect the work using normal language. Anthropic said that 90% of its code is now written this way. Few organizations will get anywhere near that figure. The idea worth taking from it is that software development becomes visible to the people requesting the work, instead of disappearing into a separate technical process.
Tableau
Tableau Studio is a rebuilt authoring experience. You can describe the visualization you want and have an agent build it, while the traditional drag-and-drop canvas remains available when you need to adjust the result manually. Skills can also be introduced during the build. One example was a skill designed to speed up a slow dashboard.
Tableau Data Apps goes beyond dashboards and lets you create an interactive application inside Claude and then publish it into Tableau, while continuing to respect the underlying permissions. There is also a data view showing the model behind the numbers on screen. That matters because one of the first questions people are going to ask about an AI-generated dashboard is, understandably, “Where did that number come from?”
Tableau Knowledge builds a knowledge graph using the semantic models you already have, along with unstructured material such as PDFs and glossaries. It can also assess the health of that knowledge and flag situations where, for example, the definition of a metric has changed without everyone realizing it.
Admin
Builder Central is probably where I would start if I were looking at the admin announcements. You describe what you are trying to build and it can create the application, agent, flow and Permission Set. Before anything is created, it gives you a brief that you can review and edit. That review step is important, because it means the workflow is not simply “prompt it and hope.” Builder Central went into beta the week after Dreamforce and is free in every Org within usage limits.
Quick Pages is aimed at a different problem. Trusted power users can create single-page applications on top of existing data without changing the underlying metadata. For all the small internal requests that are useful but never quite important enough to make it onto the development backlog, that could be a sensible middle ground.
Developer
The Salesforce development plugin for Claude Code is now generally available. It gives the coding agent secure access to a Salesforce Org, along with lifecycle tooling and pre-built skills that teach it how the Salesforce platform works. There is also a headless toolkit for other coding agents, which is important for teams that do not want their development process tied to one model or vendor.
Lifecycle Center focuses more on the everyday development workflow. You can define an orchestration, move a bug into a ready state and have an agent clone the repository, work on the fix and open a pull request. Another agent can then review the change. There is a model picker if the agent starts going in the wrong direction, along with reporting showing what work was accepted and how much time it saved. That last part will matter as organizations move past experimentation and start asking whether coding agents are making delivery faster.
Architect
The Well-Architected Framework received a substantial update, including five pillars aimed at agentic systems. There is now an architect agent in the Architecture Center grounded in that material. For a lot of people, that will be a more useful way to interact with the framework than reading through a long reference page when they are trying to answer one specific architecture question.
The Headless 360 MCP server is also intended to provide a single entry point into Salesforce using three tools: discover, describe and dispatch. It is currently in open beta and is expected to become generally available in early November as the Salesforce MCP server.
Where I Would Start
Two parts of Salesforce’s own internal story are worth paying attention to, regardless of company size.
The first is how they are organizing development work. Small squads combine an opinionated business stakeholder with a handful of builders. Stand-ups are recorded, the transcript goes into a public Slack channel, and a coding agent is brought into the process several times a day. Salesforce says that has shortened both decision-making and delivery time, and that a large share of its work will eventually operate this way. The specific process may not translate directly to every company, but the principle probably does: put the people who understand the problem close to whoever is building the solution, whether that is a person or an agent.
The second is the way they divide ownership. IT owns the foundation centrally: trust, compliance, security and the shape of the core data. Individual teams then have room to build for their function, while individual employees have room to automate or build things for their own work. That was one of the more practical answers I heard to the question of how you let more people build without creating complete chaos.
There is also a caution in all of this. The customers talking about measurable results generally did not start by buying an agent and pointing it at everything. One customer, after going through five major transformations, said the recurring mistake was not investing enough in data migration and governance. The same pattern came up elsewhere. Teams that gave agents properly catalogued, well-structured and quality-checked data reported better results and lower operating costs than teams whose agents first had to work out what the data meant.
So if I were taking one thing back from Dreamforce, it would be this. Pick one process that is worth fixing. Get the data underneath it into good shape. Put the right controls around it. Then automate or add agents from there.
That advice is not particularly new. What has changed is that most of what Salesforce announced this year makes the gap between teams that have done that groundwork and teams that have not, much more visible.
