Dreamforce 2026: The 7 Biggest Agentforce Announcements You Need to Know

Dreamforce 2026 wrapped this week in San Francisco, and if you’ve been following Salesforce’s AI story since Agentforce launched in 2024, this year felt different. Where Dreamforce ’24 was “here’s Agentforce” and ’25 was “here’s Agentforce 360,” this year Salesforce didn’t lead with one flagship product – it led with an architecture, and then hung about a dozen announcements off it.

Here’s what actually matters, cut down from the noise.

1. AIforce – The Biggest Announcement of the Week

Marc Benioff called this “the most exciting thing I’ve ever seen for Salesforce,” and it’s hard to argue with the ambition. AIforce is a live interface layer that lets agents and humans act on Salesforce data, workflows, and logic from anywhere – Slack, Claude, ChatGPT, Lightning, or a custom-built UI – without ever having to log into the traditional Salesforce screens.

The pitch: instead of a fixed UI, you get composable, intelligent interfaces wherever work already happens. Every request still runs on the requesting user’s existing Salesforce permissions, so (in theory) an agent operating through AIforce sees exactly what that person is allowed to see – nothing more.

Why it matters for admins and architects: your sharing model just became the security boundary for every new AI surface your org will ever expose. If your permission sets, sharing rules, and “temporary” access grants haven’t been audited in a while, this is the announcement that makes that overdue.

2. Koa – Salesforce’s First CRM-Native Reasoning Model

Salesforce introduced Koa, its first purpose-built reasoning model for CRM, post-trained on NVIDIA’s Nemotron 3 Super architecture. The notable detail: Salesforce says Koa was trained entirely on synthetic scenarios, with no customer data in the training corpus, and it runs inference inside Salesforce’s own infrastructure so customer data never crosses that trust boundary.

Koa is meant to handle the complex, multi-step reasoning that enterprise agent workflows need – essentially, reasoning becomes a swappable component rather than something baked into one vendor’s model.

3. The Enterprise AI Harness – The Architecture Behind Everything

This is arguably the most important announcement of the week, even though it got the least stage time. The Enterprise AI Harness is the unifying architecture Salesforce published before the keynote, spanning six capabilities: context, agency, action, governance, security, and models.

The one-line thesis, straight from Salesforce: “AI reasoning can be open-ended, but enterprise execution often cannot be.” Every other announcement this week – Koa, AIforce, the new agents, the cloud partnerships – is really just one of those six capabilities getting a product name. Once you see it this way, the whole week reads as one coherent architecture instead of a dozen disconnected features.

4. Seven New Job-Ready Agents

Salesforce shipped seven named, role-based agents built for specific jobs rather than generic “build your own agent” tooling:

• Casey – customer service across voice, SMS, WhatsApp, and chat
• Paige – IT and HR requests
• Carter – e-commerce
• Marshall – supply chain and back-office orchestration
• Piper – inbound sales qualification
• Fin – complex customer experience workflows
• Hunter – outbound sales (currently in pilot, GA expected November 2026)
• 
  

Six of the seven are generally available now. This is arguably the most concrete announcement of Dreamforce 2026 – named products with stated availability, rather than a platform capability you have to assemble yourself.

5. Headless 360 Expands – The “Sleeper” Announcement

Headless 360 turns every Salesforce cloud into a reusable enterprise capability that any authorized AI agent can discover and invoke through open standards – specifically MCP. Its MCP server lets agents running in Agentforce, Claude, ChatGPT, Cursor, and other platforms dynamically discover and call Salesforce capabilities in real time.

This one is easy to undersell, but the surface area is genuinely larger than AIforce alone – it includes tools developers may already be running against your org today, whether you know it or not.

6. Multi-Cloud, Multi-Model – Google Cloud, AWS, and Anthropic

Salesforce made a deliberate statement of cloud and model neutrality this week:

• Google Cloud: Agentforce and Gemini Enterprise connected over MCP, plus Hyperforce now running on Google Cloud
• AWS: Salesforce context flowing into Amazon Quick, AWS agents inside Slack, model choice through Amazon Bedrock, and Agentforce Voice integrated with Amazon Connect
• Claudeforce: the Anthropic partnership (announced in the run-up) that AIforce launched alongside
• NVIDIA: the foundation Koa itself is built on

Announcing Google and AWS partnerships of equal weight on the same day isn’t a Google story or an AWS story – it’s Salesforce signaling that the cloud and the model layer are both replaceable. The thing Salesforce is betting it owns is the governed business context underneath them.

7. Agent Fabric – Governance for a Multi-Vendor Agent World

MuleSoft’s Agent Fabric got expanded governance and discovery capabilities this cycle, positioning it as the control plane for agents running across Agentforce, Amazon Bedrock, Google Vertex AI, and Microsoft Copilot Studio. Its “Trusted Agent Identity” feature lets agents execute actions using a specific user’s permissions – critical once an org is running agents from more than one vendor, which is quickly becoming the norm rather than the exception.

Proof It’s Not Just Slideware: Adecco at 40+ Countries

The Adecco Group announced a global rollout of Agentforce Coworker across more than 40 countries, following a successful pilot in the UK and France. In a year full of platform announcements, this is the customer proof point – and the rollout sequence (small pilot → broad expansion) is a template worth studying for anyone planning their own Agentforce adoption.

The Takeaway

Dreamforce 2026 wasn’t dominated by one flagship product the way ’24 (Agentforce) and ’25 (Agentforce 360) were. Instead, Salesforce published an architecture – the Enterprise AI Harness – and then hung a reasoning model, an interface layer, seven agents and three major cloud partnerships off it. If you train or build on this platform, the practical shift is this: reasoning models and interfaces are now explicitly swappable, which means the durable value in any org is its data quality, business logic, and permission model – exactly the unglamorous stuff that’s easiest to defer.