Salesforce used Dreamforce 2026 to show two tightly linked moves: Koa, a reasoning model post-trained on Nvidia's Nemotron architecture for enterprise CRM work, and a new portfolio of job-ready Agentforce agents that come pre-configured for sales, service, commerce, employee experience, and back-office tasks. The company says it has already delivered 7 billion Agentic Work Units across Agentforce and Slack, including 3.2 billion in Q2 alone.
Koa: a reasoning model for enterprise work
Koa is built by post-training Nvidia's Nemotron 3 Super, a 120-billion-parameter open model. Salesforce highlighted two advantages of starting from an open base: direct access to the weights for post-training, and published training datasets that give customers a verifiable provenance story rather than a vendor black box. The model is designed to run inside the Salesforce trust boundary, with customer data and session traces staying under customer control.
The training approach is the more distinctive part. Rather than relying only on supervised fine-tuning from transcripts, Salesforce built simulated enterprise workflows across more than 14 industries, generated synthetic customers with varied moods and personas, and used reinforcement learning to score whether a customer's problem was actually resolved. The goal was to teach multi-turn tool use: carrying a task across many steps, tools, and conversation turns without losing the thread.
One simulated scenario deserves attention because it targets a fear many enterprises already have. When the right tool was unavailable, a general model tended to fudge its response—calling a similar-sounding tool that changed data it should not touch, or confirming an action that never happened. Koa was trained to respond more like a careful employee: stop, say plainly what it cannot do yet, ask for what it needs, and hand off to a human when the situation calls for judgment a model should not make alone.
A portfolio of job-ready agents
The new Agentforce agents are meant to reduce the amount of bespoke build work customers need before seeing value. Hunter is an outbound sales agent that researches prospects and collaborates with sellers. Fin is a customer agent for complex customer-experience workflows, powered by Operator and Fin Apex, custom models trained for customer experience; it is generally available now. Additional agents cover commerce, employee experience, and back-office work.
Salesforce also introduced a long-horizon runtime that lets agents pursue goals across days or weeks instead of only finishing a single interaction. Hunter is the first agent running on that runtime: a seller can ask it to rescue at-risk deals before the end of a quarter, and the agent can turn that objective into a measurable goal, build a plan, and begin working toward it while respecting the guardrails that determine when it can act autonomously and when seller approval is required. Memory preserves context across sessions, durable execution keeps plans moving as circumstances change, and dynamic steering adapts behavior based on user feedback.
Other capabilities include AI Skills for Agentforce Coworker, which lets employees teach an agent how to complete a task once and then scale that know-how across the workforce; Multi-Agent Orchestration, now generally available, for routing work across specialized agents; and Agent Optimizer, an agent that helps build, refine, test, and analyze agent performance, generally available in October 2026.
Early results and the trust question
Salesforce pointed to a mix of customer outcomes: 50% of Engine's chat inquiries resolved by its help agent Eva; 60% of Perk's sales pipeline built by Hunter; 70% of Autism Queensland's administrative requests resolved by employee service agent Paige; 90% of core shopper journeys handled by Hibbett AI, which went live in six weeks; and 4x conversation volume driven by Asana's website agent Piper, deployed in an average of 45 days. Anthropic reported that 79% of conversations Fin sees are resolved autonomously.
The trust angle is central to Koa's positioning. Because the model runs inside Salesforce's infrastructure and its traces do not train the model, Salesforce is pitching it to regulated industries—financial services, healthcare, travel, accounting—as a model optimized for enterprise work rather than a rented frontier model whose usage may become fuel for a vendor's improvement loop.
What this means
The bigger story is the shift from general assistants to specialized, governed agents that are expected to carry tasks across days, tools, and systems. Salesforce's bet is that enterprises will move faster when they can start from job-ready agents and a model purpose-built for CRM reasoning, rather than assembling everything from scratch. The unanswered questions are how much of the reported automation holds up outside early adopters, how the long-horizon runtime behaves when goals drift or tools fail, and whether the trust boundary argument is enough to win deployments in highly regulated environments.