Salesforce Agent Builder Grows Up (And Why Nonprofits Should Pay Attention)

The new Salesforce Agent Builder (beta) introduces Agent Script, a scripting language that lets you define which parts of an agent's behavior must be deterministic (same inputs grant same outputs, every time) versus which parts can use LLM flexibility. This solves the core reliability problem that made earlier Agentforce versions risky for mission-critical nonprofit workflows like benefit disbursements, donor gift management, and grant routing. A favorite quote comes to mind while reviewing this change.

“The whole function of thought is to produce habits of action.” - C.S. Peirce, “How to Make Our Ideas Clear”

After visiting a couple of "Dreamin'" events earlier this year (specifically Architect Dreamin' and CactusForce' in January), I came to the realization there's a word that has not come up enough in Salesforce Agent Builder conversations: determinism – the idea that given the same inputs, you get the same outputs.

Every time.

Not "usually."

Not "in most cases."

Every time.

It's an old concept philosophers have been arguing about since the Greeks. However, it's become newly relevant in recent times as Agentforce has continued to mature in the 16 months since it was introduced to the world at Dreamforce 2024.

To be clear on terms, Salesforce is now using Agentforce as the umbrella label for its broader AI-agent platform, which can make historical comparisons sound fuzzier than they are. In this article, when I refer to the earlier phase of Agentforce adoption, I mean the legacy Agent Builder experience; the original builder accessed through Setup, where agent behavior relied more heavily on natural-language instructions, topic configuration, and LLM interpretation. When I refer to the recent shift, I mean the new Agentforce Builder in Agentforce Studio, which introduces Agent Script and a more explicit, testable, and deterministic control layer for routing, variables, and action sequencing.

As consultants and architects have considered how this earlier Agent Builder model required us to adapt our thinking (not only about our own work, but about the work of our customers) we found that no matter how skilled one became at prompt-building, an agent’s core “attitudes” remained vulnerable to ambiguity in ways that could confirm executive fears about adopting AI strategies too early in their development.

In other words, we found that agents would decide that (somehow) this time, the user probably meant something slightly different than in other (thousands of) queries despite the prompt.

As such, many of our clients found Agentforce getting more "creative" with action sequencing than should be acceptable within the scope of trust and reliability for new operational technology (ESPECIALLY human services NPOs, to admit a bit of bias!).

And so, because of this limitation, some of the agents we had access to were making mistakes they otherwise would not make if they were configured with a more deterministic context.

This isn't to say that those who began adopting their Agentforce strategy did so in vain. Far from it! For a lot of use cases, the trade-off of earlier adoption meant these small risks were fine. However, for many nonprofits, those risks are often disqualifiers or signs of an operational opportunity not worth jumping into with confidence (yet).

After all, when a case manager logs a Benefit Disbursement, that record could affect someone's access to real services.

When a donor modifies their recurring gift, they expect the change to stick — not to be reinterpreted.

When a program officer routes a grant application, there's an audit trail that assumes the routing logic is consistent.

Nonprofits with mission critical workflows cannot often incur the luxury of being vulnerable to the subjectivity of AI prompts in a vacuum. As we have found, nonprofits have quite needed AI solutions capable of leveraging the flexibility of Agentforce’s natural language interface but only within a more disciplined, deterministic context given the stakes.

This is why the recent shift in Agentforce matters greatly! 

Salesforce spent the past year clarifying what is possible setting up Agentforce. This was not to add "features" per se, but to solve the determinism problem.

As such, the result is the new Agent Builder (now in Agentforce Studio through the ole' App Launcher waffle, not Setup) that finally enables us to specify what must happen the same way every time, and only hands off to the LLM for the parts where flexibility actually helps.

And so, the timing should be right for a lot more NPOs to consider adopting Agentforce as part of modernizing their data strategy. NPOs can now more confidently begin joining a ground floor for early adoption that has only widened since the new Agent Builder rolled into our daily reality. Nonprofits do not need an agent that is impressive most of the time. They need one that behaves when the workflow matters. That is why this version of Agent Builder is worth taking seriously. After all, inconsistency is rarely “just” inconsistency. It becomes delay, confusion, rework, and sometimes harm. A more deterministic Builder does not remove judgment from the system; it puts judgment back where it belongs.

Stay tuned for the second post in this series exploring the case for determinism. In that post I will look more closely at how the new updates solve for reliability. 

What do you think of the new builder? Have your tried it out or would you like help looking into it? Reach out to Arkus on LinkedIn or through our website contact  form with questions.