Salesforce Dreamforce 2026: Agentforce Agents, the Koa Reasoning Model and AIforce Take the Stage

Dreamforce 2026 has wrapped up in San Francisco, and if the headline numbers are anything to go by, Salesforce had no trouble filling the halls. Around 53,000 attendees showed up for this year’s edition, and the theme running through nearly every announcement wasn’t hard to spot: agents. Not chatbots bolted onto a help page, but software that’s meant to do actual jobs inside a business. Salesforce used the event to push Agentforce front and centre, introduce a reasoning model of its own called Koa, and unveil AIforce, a layer that lets outside AI tools reach into Salesforce data. You can catch up on the event itself over at the official Dreamforce page.
Seven job-ready agents, and a Coworker that’s moving fast
The centrepiece of the show was Agentforce, Salesforce’s agent platform, which now ships with seven job-ready agents aimed at common business roles. The idea is simple enough: instead of buying a blank platform and building everything yourself, you start with an agent that’s already shaped for a task and then tune it to your company.
The name that kept coming up, though, was Agentforce Coworker. Salesforce says the Coworker picked up 100,000 users in its first 35 days — a company-reported figure, so treat it as Salesforce’s own count rather than an independent audit. Still, even allowing for launch-week enthusiasm, that’s a brisk start for an enterprise tool. Enterprise software usually spreads through procurement committees and pilot programmes, not word of mouth. If a coworker-style agent is landing with individual users first, that’s a different adoption pattern than the one Salesforce built its business on.
Koa: Salesforce’s first CRM reasoning model
Underneath the agents sits Koa, Salesforce’s first CRM reasoning model. It was built with NVIDIA, and Salesforce says it will be generally available in the US in winter 2026. Reasoning models are the ones that work through a problem step by step instead of blurting out the first plausible answer, and they’ve mostly come from the big AI labs so far. Koa is interesting because it’s aimed squarely at CRM work — customer records, sales pipelines, service histories — rather than general chat.
Why does that matter? Because a generic model doesn’t know what a stalled opportunity looks like in your pipeline, or which account has a renewal coming up. A model tuned for CRM data should, in theory, reason about customers the way a good account manager does. Whether it actually does is something we’ll only learn once it’s in customers’ hands this winter, but the direction is clear: Salesforce doesn’t want to rent its intelligence from someone else forever.
AIforce opens the doors to outside AI
The announcement I found most telling was AIforce. It’s a live interface layer that exposes Salesforce data and workflows to external AI tools, so the AI your team already uses can work with your customer data instead of sitting outside it. Two examples were shown: Claudeforce, which brings that access inside Claude, and Slackforce, which does the same inside Slack.
That’s a notable shift in posture. The default enterprise play has been to lock everything inside one vendor’s walls. AIforce goes the other way — it accepts that a support lead might prefer Claude and a sales team might live in Slack, and it meets them there. It also fits a pattern we’re seeing across the industry: agents are only useful when they can reach real systems safely. AWS made a similar point this month with its local emulator that lets coding agents test cloud operations before touching production. Different product, same lesson — agent tooling is growing up, and controlled access is the price of admission.
What it adds up to
Put the three pieces together and Salesforce’s pitch at Dreamforce 2026 looks like this: agents that arrive pre-shaped for real jobs, a home-grown reasoning model that understands CRM data, and an interface layer that lets rival AI tools plug into the platform rather than fight it. The Claude connection is worth watching, too. Anthropic has been on a run lately — its Claude Fable 5.2 and Mythos 5.2 models landed earlier this month — and Claudeforce puts Claude inside Salesforce’s world instead of outside it.
Questions remain, of course. The Coworker numbers are Salesforce’s own, Koa won’t be generally available until winter, and nobody outside Salesforce has stress-tested how these agents behave on messy, real-world customer data. But the direction of travel is hard to miss. The agent race is no longer about who demos the cleverest assistant — it’s about who owns the data the agents need to do their jobs. On that front, Salesforce just reminded everyone how much of it runs through its platform.
