Salesforce AI agent governance, development key themes at Dreamforce
Salesforce goes deep under the hood of AI architecture and agentic management with new releases.
This year's mission at Dreamforce is keeping up with all the new names -- some of which represent new technologies, some previously known, and some remixes of new and old.
The high points include AIforce, a management layer and toolkit for Headless 360, and Slackforce, which combines Slackbot with a new feature called Slackforce Surfaces that, Salesforce said, will enable richer app development.
More details were released on Claudeforce, a collaboration between Salesforce and Anthropic that makes hundreds of Salesforce features available within Anthropic's Claude and enables Claude to play a more significant role in the Salesforce environment. The Claudeforce plugin will be available in public beta to joint Salesforce and Claude users Tuesday, said Patrick Stokes, president of applications and marketing at Salesforce.
Today the transition to agentic AI has begun for many frontline workers, who now click through software less and instead have agents do work for them, said Scott White, head of enterprise product at Anthropic. But it takes partnerships like Salesforce and Anthropic's to merge business understanding -- and data access guardrails -- with the automation capabilities of agentic AI. As the technology evolves, so will the partnerships.
"Agents are going to be able to take abstract business goals, really break them down into work across a variety of systems and start accomplishing higher-order business objectives," White said. "That's going to give time back and help people accomplish more ambitious things. For a seller, that's going to be building relationships with customers and, ultimately, bringing in business."
The Salesforce Enterprise AI Harness, previewed last week, is also a part of the product lineup for Dreamforce 2026, which takes place today through Thursday at Moscone Center in San Francisco. The AI harness is technology that provides data access guardrails and business logic for AI agents as they work in enterprise IT systems.
Developers, developers, developers
Another theme that promises to emerge in this week's Dreamforce conversation is the evolving role of the more than 1.5 million Salesforce developers, by some estimates, 170,000 of whom hold official Salesforce certifications. If frontline knowledge workers can build their own agentic AI automations, what will become of developers?
They will become even more valuable by concentrating on higher-level tasks and strategies, said Liz Miller, vice president and principal analyst at Constellation Research. Instead of pestering developers with service tickets, Bob the sales manager can go make his own dashboard now.
"If you're in a midsize to large enterprise and you're a Salesforce shop, you've really been grinding on your own," Miller said. "Now, all of a sudden, you become creative. Now you're building the business applications. Now you're building the business interfaces. Now you're building the experiences. Devs are suddenly able to focus on how the business runs on technology, rather than Bob down in sales who can't get his projections right."
What all this means for users
Many of this year's Dreamforce announcements are tied to security, infrastructure and governance -- much less shiny objects than the unveiling of Agentforce two years ago. But Salesforce needs to convince CIOs and tech buyers that it's safe to use after they've heard two years' worth of stories about rogue agents and doomsday scenarios, and it will take more than simply rattling off a list of the platform's technical capabilities.
Continuously building trust with CIOs begins with understanding how the CIO's job description is evolving along with agentic AI, said Rohan Kumar, president and chief platform and engineering officer at Salesforce. CIOs now must deal with the coming agent sprawl and protecting business assets, and companies like Salesforce must design systems such as its Enterprise AI Harness to keep them working.
"Without having good guardrails, it's impossible to transition to [the agentic] enterprise," Kumar said. "I think the first step for us is deeply understanding and appreciating the challenge we need to go solve, and that's the journey we are on."
Really, the problem with rogue AI has nothing to do with it becoming sentient or overstepping its guardrails, Miller said. Like all software, she reminded us, AI comprises ones and zeroes doing what they are instructed to do. The governance piece will be where the mistakes are made and prevented.
"The thing that Salesforce has very intentionally done [since the earliest days of Einstein AI] is to talk about a governance layer that needed to not only understand your data and your systems but understand your workflows and processes," she said. "Salesforce understands that verification, validation, access and authorization are discrete, specific and all critically important."
Also discussed by Salesforce execs will be Koa, a newly named CRM-based foundational large language model that Salesforce has had in development for years and has recently enlisted Nvidia to bring across the finish line.
Koa will help customer service and sales reps, for example, determine the next best steps to complete deals or close cases, Kumar said. Few other details were on offer in a Monday preview to media and analysts, but more is likely to come over the next week.
Don Fluckinger is a seasoned B2B technology journalist with more than 30 years of experience specializing in enterprise IT, digital experience and content management. As a senior news writer at Informa TechTarget, he delivers award-winning analysis that helps IT and business leaders navigate complex technologies to enhance customer and employee experiences. Got a tip? Email him.
