Salesforce’s AI Frontier Is Long-Horizon Work
By Supreeth Koundinya ·
At Dreamforce 2026, Salesforce defined what frontier means for a CRM (customer relationship management) company in the agentic age. It’s all about its AI platform Agentforce’s new long-horizon runtime, multi-agent orchestration, and its growing portfolio of job-specific agents.
It is trying to move CRM agents beyond unidimensional tasks to pursuing goals over days or weeks, retaining memory, adapting to new information, and handing work to other agents when required.
Salesforce says its new runtime now gives agents memory, durable execution, and dynamic steering, and its new agent ‘Hunter’ is the first agent built on it.
Hunter is designed for outbound sales, involving researching prospects, initiating outreach, managing responses, and adjusting its plan as new information arrives. The agent can keep working toward a sales objective after the original interaction ends, while keeping the seller in control.
Other agents in the portfolio apply the same direction to different business functions.
Casey is a customer-service agent who handles tasks such as answering questions, processing returns, and managing account requests.
Paige handles IT and HR service requests, while Carter assists shoppers with product discovery, comparisons, and checkout.
Marshall focuses on supply-chain and back-office processes, Piper works on inbound lead engagement and qualification, and Fin is a customer-service agent.
Salesforce says these agents are designed to give customers a starting point for deploying specialised AI workers rather than requiring every workflow to be built from scratch. Six of the seven agents are generally available; Hunter remains in pilot, with general availability planned for November.
“All of these agents are built on top of the Agentforce stack—which is our harness that brings together the family of models, both internal and external data from Data Cloud and Data 360,” Jayesh Govindarajan, the company’s VP of AI, tells AIM.
Building a Durable Harness
With long-horizon tasks, Govindarajan says harnesses need to be more durable and deeply understand goals. “It needs to keep track of the first task that was partially done as the second task gets kicked off—and then go back to finish the first task, and so on.”
Salesforce’s new runtime is designed to do that through three capabilities: memory, durable execution, and dynamic steering.
Memory preserves context and progress across sessions, while durable execution keeps a plan running over time and lets an agent resume or course-correct when circumstances change. In addition, dynamic steering lets a user modify the agent’s behaviour while the work is underway.
The company’s example is an account executive asking Hunter to “re-engage my at-risk deals.”
Hunter can examine open opportunities and engagement signals from Data 360 and Slack, turn the request into a measurable objective, such as re-engaging a certain portion of an at-risk pipeline worth hundreds of thousands of dollars, and formulate a plan spanning days, weeks, or months.
That plan includes the tasks to be performed, their timing, the context required, and the guardrails determining when Hunter can act autonomously and when it must seek seller approval.
Salesforce has spent the past year breaking down agentic tasks into smaller, specialised subtasks, including understanding intent, selecting the right tool, and finding the relevant context. Rather than relying on a single frontier model to perform all of these functions, Salesforce has built smaller models for specific parts of the workflow.
This is tied to Salesforce’s broader agentic stack.
Data 360 supplies the context layer, bringing together internal Salesforce data and external data connected to the platform. The action layer provides the APIs and workflows that execute work, while the Agentforce harness determines which model, context, and action should be used for a particular task.
The company is building on Agentforce’s mass appeal. During the conference, Salesforce revealed that Agentforce and Slack delivered 7 billion Agentic Work Units, including 3.2 billion in the second quarter alone.
Making Long Horizon Tasks Cost Efficient and Accurate
Govindarajan says Salesforce’s understanding of CRM workflows allows it to specialise models for the tasks customers perform most frequently.
A model handling an order-status request, for example, does not need the same capabilities as one working through a complex account strategy. The system can route each task to the right model, while retaining the option to use a frontier model for more difficult reasoning.
“We rely less and less on a common frontier model to do it all,” he says.
Smaller models can reduce latency and inference costs for routine CRM tasks, while the company’s own reasoning models can be tuned for enterprise-specific work
Besides, Salesforce’s approach to reducing hallucinations begins at the individual-agent level. Govindarajan explains that developers can use Agentforce’s agent scripting language to specify how an agent should behave and where it should rely on deterministic actions rather than model-generated responses.
For open-ended tasks, an agent may need to reason, interpret a request or decide what to do next. But when a required operation already exists as an API, the agent should call that API instead of generating code or improvising an answer.
The same scripting mechanism is used when multiple agents are involved.
However, multi-agent systems introduce another requirement: discovery. An agent must identify which other agent is appropriate for a particular task before handing work over.
That requires each agent to register high-quality metadata describing its capabilities, so other agents can identify and select it within the agent network. Once the system finds the right agent, it needs a defined communication layer. Salesforce uses the agent-to-agent, or A2A, communication protocol for this exchange.
What is Salesforce’s Koa?
Salesforce had previously relied on general-purpose models for the core reasoning required to complete more complex workflows. Now, it has released Koa, its first CRM reasoning model, developed with NVIDIA and built on the open-weight NVIDIA Nemotron 3 Super 120B model.
Salesforce post-trained Nemotron using synthetic scenarios modelled on nearly three decades of CRM deployments, covering tasks such as updating opportunities, routing service cases, and scheduling follow-ups. The company says Koa matches or exceeds leading model performance on its CRM benchmark while producing 3x fewer errors.
“Frontier models use the same base model for solving a refund problem as they do for solving a physics problem,” Govindarajan notes. “Our point of view is that you should find the right tool for the right job, and the agentic system should be the harness that enables that choice.”
The model is designed to handle a broad range of work, from simple tasks such as checking an order status to complex workflows such as optimising a contact centre. Salesforce’s answer is a routing layer that selects models according to the task.
Koa is therefore designed for CRM-specific reasoning rather than general intelligence. Salesforce says it offers three advantages: lower cost and latency than using a frontier model for every task, the ability to run the model within
Salesforce’s own infrastructure, and greater accuracy on CRM workflows.
Salesforce controls the model weights, while post-training and inference occur within the company’s trust boundary. It is now a part of a wider family of models inside Agentforce.
Govindarajan says smaller language models handle tasks such as intent detection, tool calling, and context selection. They are designed to be low-latency and token-efficient, while more demanding tasks can still be routed to frontier models.