Dreamforce 2026 Keynote Recap: AIforce, Koa, and Seven Named AI Agents
Dreamforce 2026 opened this week in San Francisco with Marc Benioff tackling the "SaaSpocalypse" narrative head-on: this isn't the end of software, he argued — it's possibly the end of software that makes humans do all the work.
His framing lines up almost exactly with what we laid out in Agentforce vs. Claudeforce: You Don't Have to Pick a Side: enterprise interfaces have moved from command line to GUI to web to mobile, and AI is simply the next stage of that same evolution. This week's keynote filled in a lot of the specifics — a new interface layer, Salesforce's own reasoning model, and seven named AI agents with actual job titles.
Here's what actually shipped, what's still a pilot, and what it means for teams already running Agentforce.
The Keynote Framing
Dario Amodei joined Benioff on stage and compared today's AI market to the early motor industry: even if model progress froze completely right now, he argued, most enterprises are only using something like 5–10% of the value already sitting in the technology they have access to.
Four New Pillars
Four names carried the keynote. Each one answers a different question about where AI sits relative to Salesforce's existing platform.
AIforce — the live interface layer
AIforce carries Salesforce's data, workflows, business logic, permissions, and governance into whatever AI surface a team is actually using. It launched alongside Claudeforce, Slackforce, and Agentforce Coworker. Its AgentExchange partner ecosystem already spans interfaces from Anthropic, AWS, and Google, plus tools from Docusign, Gamma, Jasper, Ramp, and Rippling.
Benioff repeated one point often: zero data retention. Customer data doesn't train models outside a customer's own trust boundary, regardless of which AI surface it flows through.
Claudeforce — Claude inside the stack
The partnership we covered last time got its keynote moment, with Benioff's "number one AI meets the number one CRM" line and Amodei's early-motor-industry analogy for how much headroom is still unused.
Slackforce — an AI layer built into Slack
A new intelligent AI layer inside Slack, paired with "Slack Code" for building things collaboratively without leaving the app.
Koa — Salesforce's own CRM reasoning model
Arguably the biggest surprise of the day. Koa is built by post-training NVIDIA's Nemotron 3 Super on a synthetic dataset modeled on nearly three decades of CRM deployment knowledge — no customer data used in training. Salesforce says it matches or beats leading general-purpose models on CRM actions, with three times fewer errors.
Pilot now · GA winter 2026
The Momentum Numbers
Benioff leaned on two figures to make the case that this isn't early-stage adoption anymore:
Seven Named Agents, Each With a Job Title
Salesforce pre-announced seven "job-ready" agents on September 11 to set up the keynote — each with a name, a specific job, and a status rather than a vague capability list.
| Agent | Job | Status |
|---|---|---|
| Casey | Customer service — voice, SMS, WhatsApp, web chat | GA now |
| Paige | Employee IT & HR requests | GA now |
| Carter | Commerce — product discovery, in-chat checkout | GA now |
| Marshall | Supply chain & back-office ops | GA now |
| Piper | Inbound lead qualification & conversion | GA now |
| Fin | Customer experience workflows | GA now |
| Hunter | Outbound sales — works a pipeline over weeks | Pilot, GA Nov 2026 |
Two are worth flagging specifically. Hunter runs on a new "long-horizon runtime" built to keep pursuing a goal over days or weeks instead of resetting after each session — a meaningful shift from the single-session agent pattern most Agentforce deployments use today. Fin, recently acquired by Salesforce, is already autonomously resolving 79% of Anthropic's own support conversations — a genuinely strong reference case for a customer-experience agent.
Agent Script and the Governance Layer
Underneath all seven named agents, Salesforce introduced Agent Script — an open-source language that blends AI reasoning with deterministic rules, specifically aimed at cutting down on unpredictable agent actions. It's the same instinct we flagged in our last post: separate what the model decides from what the platform enforces, and do it in code, not just in policy.
Alongside it, Salesforce introduced the Trusted Enterprise AI Harness — a governance layer purpose-built for enterprises running multiple agent platforms at once, not just Salesforce's own.
This is the governance question we said was worth auditing before Dreamforce even happened: Agent Script and the Trusted Enterprise AI Harness are Salesforce's answer to "how do we keep agent behavior predictable across platforms." Whether it's sufficient for your compliance requirements is still worth evaluating on its own.
Agentforce Coworker: The Orchestrator
Agentforce Coworker isn't a single named agent so much as the orchestrator sitting above the others. It reasons across accounts, activity, and history, and calls whichever specialized Agentforce agents an organization has already deployed — Casey, Paige, Piper, and the rest — rather than requiring a user to know which agent to invoke.
New this week, it ships with AI Skills: an employee teaches it a task once, and that know-how scales across the workforce. It's the clearest sign yet that Salesforce is thinking about agents less as point tools and more as a workforce that shares what it learns.
What This Means for You
This keynote doesn't change the core architecture question from our last post — it sharpens it. There are now two reasoning options sitting inside the same governed stack: Claude (via Claudeforce) and Koa, Salesforce's own CRM-tuned model. AIforce formalizes what we called the "reasoning layer" into an actual named product with its own partner ecosystem. And Agent Script plus the Trusted Enterprise AI Harness are Salesforce's first real answer to the "governed action layer" half of the equation.
Two Reasoning Options
Claude and Koa now both sit in the reasoning layer. Which one you route a workflow to becomes a real design decision, not a hypothetical one.
Governance Just Got a Name
Agent Script and the Trusted Enterprise AI Harness are worth evaluating against whatever governance approach you already built.
Named Agents, Real Jobs
Seven agents with job titles is a different adoption story than "an AI platform" — easier to scope, pilot, and measure against a single role.
None of this changes the audit we recommended last time. If anything, it raises the stakes: the sooner your Agentforce build cleanly separates what a model decides from what the platform enforces, the easier it is to slot in Koa, Claude, or whatever reasoning layer wins next — without a rebuild.
Dreamforce 2026 Keynote: FAQ
What is AIforce?
At Dreamforce 2026 (September 15–17, San Francisco), Salesforce launched AIforce, a live interface layer that brings Salesforce data, workflows, business logic and permissions into the AI tools people already use. Its first surfaces are Claudeforce inside Claude, Slackforce inside Slack and Agentforce Coworker inside Salesforce.
What is Koa?
Koa is Salesforce’s first CRM reasoning model, announced on September 15, 2026. Salesforce built it by post-training NVIDIA’s Nemotron 3 Super on synthetic CRM scenarios and says no customer data was used to train it. It was in pilot at launch, with general availability expected in winter 2026.
Which named Agentforce agents did Salesforce announce?
On September 11, 2026, ahead of Dreamforce, Salesforce announced seven job-ready agents: Casey (customer service), Paige (IT and HR requests), Carter (commerce), Marshall (supply chain and back office), Piper (inbound leads), Fin (customer experience) and Hunter (outbound sales). Six were generally available at announcement; Hunter was in pilot, with general availability planned for November 2026.
What are Agent Script and the Trusted Enterprise AI Harness?
Agent Script is Salesforce’s open-source language for agent behavior, which mixes AI reasoning with deterministic rules. The Trusted Enterprise AI Harness, introduced in September 2026, is a shared architecture for running and governing agents across Salesforce and third-party AI. Both aim to make agent actions more predictable, but they still need checking against your own compliance requirements.
What do the Dreamforce 2026 announcements mean for teams already on Agentforce?
The core advice is the same: keep what the AI model decides separate from what the platform enforces. With Claude and Koa both available as reasoning options, a build that makes that split cleanly is easier to move between models without a rebuild.
The Bottom Line
Dreamforce 2026 turned "the agentic enterprise" from a keynote theme into a named product lineup: AIforce, Claudeforce, Slackforce, Koa, seven job-titled agents, and a governance layer to hold it all together.
The fundamentals haven't changed. What's changed is how much of the stack now has a name — which makes it a much better time to check whether your own Salesforce architecture is actually ready for it.
Ready to see where your Salesforce org stands?
Schedule a free Agentforce Readiness Assessment with Brazen Group. We'll evaluate your current build against the reasoning-layer / governed-action-layer split and give you a clear roadmap for AIforce, Koa, and whatever ships next.
Related reading: Agentforce vs. Claudeforce: You Don't Have to Pick a Side