In this post, we take a deep dive into the new Agentforce Builder and walk through what has changed, what it means for developers, and how it affects the way agents are built and deployed.
The previous builder relied almost entirely on LLM interpretation. You wrote instructions and the Atlas Reasoning Engine decided what to do. For simple use cases that works well. For production workflows where specific rules need to execute the same way every time, it created unpredictable behavior that was difficult to debug and harder to govern.The new builder addresses this directly with Agent Script, a new scripting language for agent logic, alongside a rebuilt workspace, in-builder testing, and a VS Code integration that brings agents into your existing deployment pipeline.
What is the New Agentforce Builder?
The new Agentforce Builder is a completely rebuilt interface for creating and managing agents. The biggest structural change is that everything now lives in a single workspace. The old builder split Topics, Actions, Testing and Settings across separate tabs, which meant constant context switching.The Explorer sidebar now gives you a flat, navigable tree of your entire agent including subagents, actions, variables and connections, all visible at once.
Here is what the same agent looks like in the new builder:
Canvas View and Script View
The editing area supports two views that stay in sync with each other. Canvas View renders agent configuration as structured visual blocks, which is useful for reviewing logic quickly or for admins making changes. Script View is where developers write Agent Script directly, with syntax highlighting, autocompletion and real-time validation.
Think of it like a spreadsheet with a formula bar. Canvas View shows you the formatted result, Script View shows you the underlying formula. Editing either one updates the other automatically.
What is Agent Script?
Agent Script is a new scripting language for defining agent logic. It sits between natural language instructions and Apex code on the complexity scale.
Before Agent Script, the agent was essentially a black box. You wrote instructions and trusted the LLM to apply them correctly. Agent Script changes that by letting you define the logic explicitly.
With Agent Script you can:
- Define conditions using if/else logic based on context variables
- Chain actions in a guaranteed sequence before the LLM takes over
- Manage variables across subagents
- Control transitions between subagents explicitly
The underlying reasoning engine has also shifted to a hybrid model. Deterministic paths are defined in Agent Script. Conversational parts remain LLM-driven. You control which parts are which.
How Does Testing Work?
The Preview panel is built directly into the workspace. You no longer need to switch tabs or leave the builder to test a change.
Live Test Mode lets you set context variables, run a conversation using mock data and see a reasoning trace per message showing which subagent was selected, which actions fired and why.
Here is an example where a customer asked to create a case for a login issue. The Interaction Summary on the right shows each reasoning step the agent took:
You can also drill into the Trace view for a granular breakdown of every step with millisecond timings:
Agentforce Grid is a bulk testing environment available under Setup > Testing Center. It lets you run a test suite across multiple utterances at once and review the results in a single view.
What is the Agentforce Assistant?
The Agentforce Assistant is an AI panel embedded in the builder. You describe what you want in plain English and it generates subagents, actions and Agent Script for you to review and accept.
It is particularly useful for generating test utterances, which would otherwise need to be written manually:
What is Agentforce DX?
Agentforce DX is a developer toolchain that lets you work with agents outside the browser. You can pull agent definitions including Agent Script files into a local Salesforce DX project, edit in VS Code with full language support, and deploy using standard sf deploy commands.
Agent Script is version-controlled metadata. That means pull requests, code reviews and rollbacks apply to agent logic the same way they apply to Apex classes and other org metadata.
Key Takeaways
- Agent Script replaces prompt-only instructions with deterministic, programmable logic
- The hybrid reasoning engine lets you mix guaranteed paths and LLM-driven paths in the same agent
- Live Test Mode and Agentforce Grid make agent behavior verifiable before deployment
Agentforce DX brings agents into source control and your existing CI/CD pipeline
