You don't have to leave it all to the LLM—Agentforce design changed by Agent Script

The other day, I was invited to Salesforce Tower as a Community Leader to attend the Agentforce Press and Analyst Briefing!

In this article, I will summarize my thoughts, focusing on the design philosophy of Agent Script, which I found most impressive.

First, starting from where Agentforce is now.

At the beginning of the presentation, several figures were shared.

In about a year and a half since its release, there are over 18,500 contracts, a 330% increase year-over-year.
Tokens issued to external LLMs are over 3.2 trillion.

Honestly, this scale is amazing. And Salesforce positioned this as the basis for declaring the "experimental phase is over."
It is a declaration that agent utilization is moving from the "try it out" phase to the "deliver results" phase.

The problem of not being able to reach production

On the other hand, one thing that was impressive during the presentation was this statement:
" Many customers say they can't reach production even after starting with agents"

Facing this challenge head-on at their own product announcement event.
I felt a sense of sincerity in that, and also empathy, thinking, "Ah, I thought so."

Two main reasons were cited for not being able to reach production.

① Data is not reaching / Quality is insufficient
Because reliable context cannot be created, agents do not function properly.

② AI behavior appears as a black box
The anxiety of "not knowing what it will do" makes people hesitate to deploy it to production.

I think these two are walls that you hit with any AI implementation, not just Agentforce.

What Agent Script is trying to solve

This is the main point.

LLMs (Large Language Models) are good at interpreting natural language.
However, because of that, there is an aspect where it is difficult to make them follow procedures that must never be missed, such as "always check the email address first."

Even if you instruct them in natural language to "please check the email address first," the agent might forget to check it.
Even if you write complex prompts, instability remains. This is also a structural problem where the "flexibility" and "certainty" of LLMs are prone to being a trade-off.

Agent Script is a scripting language designed to resolve this trade-off.

While strictly defining deterministic logic like if/else in the Script, you leave the parts that require natural language processing to the LLM.
It is the idea of a hybrid of "flexibility (LLM)" and "certainty (rules)".

For example, in critical business processes where procedures must be strictly followed such as refund processing or loan applications, Salesforce's answer to that reality is Agent Script.

What changes with the new Agentforce Builder

Agent Script can be used on the updated Agentforce Builder.

The interface has been refreshed, and it is designed to allow you to switch between a Canvas screen where natural language and scripts coexist, and a Script-only screen where only the script is displayed.

Personally, what I found exciting is the feature that allows even non-developers to generate scripts. AI assists in converting natural language into scripts. It seems like we will be able to discuss Agent design more as a team.

Also, the fact that you can duplicate an Agent simply by copying the script is great when you consider deployment speed in practice!

Intelligent Context in Data360

Another announcement is the enhancement of Intelligent Context.

By simply uploading documents such as PDFs or manuals that include diagrams and tables, you can use them as sources for Agent responses. It is said to be configurable with no-code and easily set up even by an Agent.

The important premise is that without 'trusted context,' an agent cannot function no matter how smart it is. Even if you make the behavior deterministic with Agent Script, if the quality of the referenced data or documents is low, the output cannot be trusted. Intelligent Context is a feature that directly addresses that data quality issue.

The reality shown by implementation examples

During the announcement, an in-house implementation example from a certain company was also introduced. (I will omit the details as they are not for public release!)

They built Version 1 in about 2.5 months, and they are already seeing results. However, they mentioned that
'human checks are still necessary for operations'.

And there was another impressive figure. It was a case where
preparation time before talking to customers was reduced from 40 minutes to 40 seconds.

What I want to highlight here is that the Agent is not 'standing in front of the customer,' but rather playing an active role by supporting preparation before a human handles the interaction. The Agent takes on the task of collecting and organizing relevant information, and the human reviews that content before talking to the customer. I felt this was exactly how to use it as an
Employee Agent, and a textbook example of Human-in-the-loop.

We tend to have the image that 'AI Agent = something that interacts directly with customers,' but a design where the human is in front while the Agent supports from behind is a very realistic approach from the perspectives of both reliability and ease of implementation.

If you run with the expectation that 'AI will do everything!', it surely won't go well. I felt that the approach of
starting small and expanding the scope while incorporating human checks fits the current reality.

What I thought was most important

After listening to the announcement, what I felt again can be summarized as follows.

When people say 'AI' or 'AI Agent,' it's easy to think they can do anything from the start. But in reality, that's not the case.

Just as you need to train people when you hire them, agents won't work perfectly from the start. It is essential to understand the process of gradually expanding the scope of what you entrust to them
.

Agent Script is a tool that allows designers to properly control that "scope of entrustment." Instead of leaving everything to the LLM, you can use scripts to guarantee the parts that "must be done this way," while leaving the rest to the flexibility of the LLM.

Perhaps you could call it the feeling of raising an AI.

In the agent design you are currently working on, are the "parts to entrust" and the "parts that cannot be compromised" clearly separated?

Agent Script seems like it will be a new way to face that question!