Beyond Conversations: Understanding Long-Horizon Agents in Salesforce Agentforce

Artificial Intelligence has changed how businesses interact with customers. From answering customer questions to automating repetitive tasks, AI agents are becoming an important part of modern business applications.

With Salesforce Agentforce, businesses can build AI agents that understand customer requests, access CRM data, and perform actions using Salesforce tools such as Flows, Apex, and external integrations.

But there is an interesting question:

Imagine a company managing hundreds of sales opportunities in Salesforce. Some customers are actively engaging with sales representatives, while others have stopped responding after initial discussions.

A sales opportunity may involve multiple activities, such as understanding customer requirements, scheduling product demonstrations, sharing proposals, negotiating pricing, coordinating approvals, and following up with decision-makers.
This process may take several days or even weeks.

A traditional conversational interaction is primarily focused on responding to the current request. However, a complex business process requires something more: the ability to remember previous interactions, maintain execution progress, handle interruptions, and adapt when requirements change.

This is where Long-Horizon Agents come into the picture.

Instead of focusing only on individual conversations, Long-Horizon Agents are designed to pursue goals across extended periods, maintaining continuity throughout complex workflows.

In this article, we’ll explore what Long-Horizon Agents are, understand their three important capabilities — Memory, Durable Execution, and Dynamic Steering — and see how they can transform real-world Salesforce business processes.

Before understanding Long-Horizon Agents, let’s look at how a traditional AI agent generally works.

A typical conversational interaction follows this pattern:

For example, a sales representative asks:”Can you show me the current status of my open opportunities?”

The Agentforce agent retrieves the relevant Opportunity information from Salesforce and responds.
This works perfectly for simple requests.

However, consider a different request:

“Help me re-engage at-risk opportunities and improve our chances of closing them this quarter.”

This is not a single action. It involves multiple activities, decisions, follow-ups, and potentially several interruptions.

A Long-Horizon Agent is designed for this type of work.

It can:

Salesforce introduced a Long-Horizon Runtime for Agentforce to support agents that pursue goals over days and weeks instead of completing only an individual task or interaction.

One example is Hunter, Salesforce’s outbound sales agent.

A sales representative can ask Hunter to re-engage at-risk opportunities before the end of a quarter.

Instead of simply generating an email, Hunter can analyze the sales pipeline, establish a goal, create an outreach plan, execute activities over time, and adjust its approach when circumstances change.

This represents an important shift in AI development:

From responding to requests to taking responsibility for pursuing business goals.

Long-Horizon Agents rely on three important capabilities:

It is important to understand that Salesforce CRM data and agent memory are not exactly the same thing.

CRM records store structured business information. Agent memory helps preserve relevant context and progress that the agent needs when continuing its work.

For example:

The agent can combine relevant business data with remembered context to make better decisions.

Key takeaway: Memory reduces repeated conversations and helps an agent maintain continuity across multiple interactions.

Memory helps an agent remember information. But remembering information alone is not enough.An agent also needs to know what it was doing, what has already been completed, and what needs to happen next.

This is where Durable Execution becomes important.

Think of it like saving your progress in a video game.

Imagine you have completed five levels. Suddenly, your system shuts down. When you return, you don’t want to start from level one.

You want to continue from where you stopped.

Durable Execution follows a similar idea for long-running agent tasks.

Let’s continue with our sales opportunity example.

The agent has the following execution plan:

Suppose the agent completes the first three steps, but the customer postpones the product demonstration.

The agent should not unnecessarily repeat all the completed activities.

It needs to preserve execution progress and continue when the next action becomes appropriate.

The agent can use the relevant Salesforce records and its execution state to determine what has already happened and what needs attention.

It can also use Salesforce actions, such as Flow, Apex, and external integrations, to perform specific business operations.

For example, if the customer postpones the proposal discussion, the agent can adjust the next follow-up activity rather than restarting the entire sales process.

Important distinction: Durable Execution is not simply another name for a Scheduled Flow or a Record-Triggered Flow.

Salesforce Flows and Apex can already automate scheduled and event-driven operations. Durable Execution addresses the additional challenge of preserving an agent’s ongoing plan and execution progress across a long-running task.

Key takeaway: Durable Execution helps an agent continue working toward its goal without unnecessarily restarting the entire process.

Business processes rarely follow a perfectly defined path.

Customers change their minds. Employees provide new instructions. Opportunities become more or less important. Unexpected situations occur.

An AI agent must be able to adapt.

This is where Dynamic Steering comes into the picture.

Let’s continue with our sales opportunity example.

The customer initially agrees to purchase a software solution.

The agent creates a plan:

However, after reviewing the proposal, the customer says:

“Our budget has changed. Can you suggest a different package within our revised budget?”

The original plan is no longer completely relevant.

The agent needs to adjust its approach based on the new instruction.

Another example comes from Salesforce’s Hunter sales agent.

Suppose Hunter is working to re-engage an at-risk opportunity. It sends an email, but the email bounces because the contact is out of the office.

Instead of blindly continuing with the original plan, Hunter can investigate an alternative contact, prepare a revised outreach message, and request approval when the change requires it.

This is an example of adapting the execution plan based on new circumstances.

It is important to remember that adapting a plan does not mean ignoring business rules.

For example, an agent should not automatically approve a discount or modify commercial terms simply because a customer requests it.

Business rules, permissions, approval requirements, and human oversight must still be respected.

Key takeaway: Dynamic Steering allows an agent to adjust its behavior as business requirements and circumstances evolve.

Agents can respond to changing customer requirements and unexpected situations.

Salesforce’s introduction of the Long-Horizon Runtime represents an important step in the evolution of Agentforce.

Hunter is an example of how AI can work toward sales goals over extended periods.

Salesforce has also indicated that customers will be able to build and customize their own Long-Horizon Agents using Agentforce as the platform evolves.

This creates interesting opportunities for Salesforce developers.

Imagine future implementations where agents can:

These are examples of potential applications, not claims that every use case is already available as an out-of-the-box capability.

For Salesforce developers, the focus will increasingly move beyond building agents that simply answer questions.

It will involve designing agents that understand business objectives, interact with enterprise systems, maintain continuity, and operate within defined business boundaries.

AI agents are evolving from conversational assistants into systems that can pursue meaningful business objectives over time.

Long-Horizon Agents address an important limitation of traditional conversational AI: the difficulty of managing complex work that extends beyond a single interaction.

With Memory, agents can maintain relevant context.

With Durable Execution, they can preserve their plans and continue working.

With Dynamic Steering, they can adapt when circumstances or instructions change.

When combined with Salesforce CRM data, Data 360, Flow, Apex, and enterprise security controls, these capabilities open new possibilities for building intelligent business processes.

The future of AI agents is not just about answering the next question. It is about making meaningful progress toward the bigger goal.