Loading

Learn About New Agentforce Builder with Agent Script

Udgivelsesdato: Mar 2, 2026
Beskrivelse

What is Agentforce Builder with Agent Script?

Agentforce Builder, also known as Next Gen Authoring, is Salesforce's next-generation agent authoring platform that empowers both low-code and pro-code developers to build intelligent agents with unprecedented control and reliability. The platform introduces Agent Script, a revolutionary programming language that enables hybrid reasoning, seamlessly blending the creativity of LLMs with deterministic business logic.

Key Components:

  • New Builder UI: Completely redesigned interface with conversational, low-code, and pro-code authoring modes
  • Agent Script: New scripting language combining natural language instructions with programmatic expressions
  • Hybrid Reasoning: First-of-its-kind control over agent behavior, opening up the Atlas reasoning engine

Latest Release Features:

  • Variable Management UI: Originally, variable management was Script only. Now customers have the option to use a low-code, simplified UI.
  • ASA and AEA Template Support:  Both ASA (Agent Service Agent) and AEA (Agentforce Employee Agent) templates are supported in the new Agentforce Builder. 
  • Self-Healing:  Agent understands prompt context and auto-recovers from UI changes, repair variable mappings, adapt to configuration changes 
  • Connection Parity:  All existing connections or channels supported in the legacy builder are fully supported in NGA
Løsning

Implementation Strategy

Phase 1: Foundation 

Recommended Approach: Conversational Mode

  • Use the embedded Agentforce assistant to describe business requirements
  • Let the assistant generate initial topics, actions, and instructions
  • Focus on getting basic Agent structure in place
  • Recreate an already functional Agent using the conversational interface

Best Practice: Follow the 5 Levels of Determinism Framework

  • Level 1: Structure topics and actions to avoid semantic overlap
  • Level 2: Write clear, guiding instructions without over-scripting
  • Level 3: Use RAG for grounded, factual responses
  • Level 4: Implement variables and state management for complex interactions
  • Level 5: Hand off to deterministic processes (Apex, APIs, Flows) when needed

 

Phase 2: Enhancement 

Move to Canvas View:

  • Use / command for quick if-then patterns
  • Use @ command to insert variables and resources
  • Add deterministic transitions between topics based on specific conditions

Testing Best Practices:

  • Use Simulate mode (mock data) for early testing to avoid org changes
    • Note - Simulate mode will consume credits
  • Progress to Live Test when ready for actual data interactions
  • Leverage Agentforce Testing Center with your prepared question set

 

Phase 3: Advanced Logic 

Transition to Script View (for technical or pro-code users):

  • Implement complex business rules using Agent Script expressions
  • Add action chaining for sequential workflows
  • Create conditional logic for branching scenarios

 

Success Factors

Agent Architecture:

  • Create Global Instructions Topic for universal guidelines (tone, escalation rules)
  • Implement Fallback Topic to handle unmatched requests with knowledge search
  • Use custom search index and retriever instead of default Data Library auto-generated ones

Testing & Quality:

  • Test every agent version against your question set using Testing Center
  • Use Interaction Details and Trace for debugging agent decisions
  • Monitor variable values throughout conversations to understand agent reasoning

 

Frequently Asked Questions

QuestionAnswer
When is Agentforce Builder with Agent Script available?
  • February 2026
Does the new builder cost extra?No additional cost if you have purchased Agentforce. The new builder is available at no charge in Agentforce Studio.
Do Agents created by the new Builder consume credits (or remain unmetered) in the same way as the previous Builder?Yes
What's the difference between Canvas and Script editors?Both views maintain consistent agent configuration:
  • Canvas: Visual blocks with shortcuts for low/no-code users (/ for patterns, @ for resources)
  • Script: Direct code editor for precise control with syntax highlighting and validation
  • Conversational: Chat with Agentforce assistant to describe requirements in natural language
Can I migrate existing agents immediately?Not during open beta. You must create new agents from scratch. Cross-over options will be considered closer to GA to allow for ease of setup. 
What happens to my current agents?Existing agents remain fully functional and unaffected. Both builders coexist in the same org.
Can I have both old and new agents in the same org? Yes! Both builders work simultaneously.
Are all the same action types supported? Yes - Apex classes, flows, external APIs, and prompt templates are all supported, same as legacy builder.
How do I prevent common mistakes? The builder includes real-time validation with immediate feedback:
  • Critical Issues: Syntax errors, undefined references, missing variables
  • Warnings: Unused variables, data type mismatches, unreachable logic
  • Proactive Hints: Suggestions to convert complex instructions to structured expressions
Can multiple users edit an agent simultaneously? Technically possible but not recommended. Changes aren't visible until saved, creating risk of conflicts. Check "Modified By" column to see recent changes.
What's the difference between Simulate and Live Test? 
  • Simulate: Uses mock data generated by LLM, no org changes, relaxed validations for early testing
  • Live Test: Runs actual actions, makes real org changes, same as legacy preview but with enhanced tracing
How can I see why my agent made specific decisions? Use Interaction Details for high-level AI-generated explanations of topic/action selection, and Trace for detailed step-by-step spans. Check Variables tab to see values influencing decisions.
Can I test voice agents? Voice testing is coming soon. Voice agents can be created and deployed through the Connections tab.
What can the Agentforce assistant do? The embedded assistant helps with:
  • Converting natural language to valid Agent Script
  • Creating agents from scratch with clarifying questions
  • Analyzing errors and providing fix suggestions
  • Debugging trace data and troubleshooting issues
Does using the assistant consume credits? No, the Agentforce assistant is not billed or metered and doesn't consume credits.
Can the assistant make changes without approval? No, the assistant cannot make functional changes without explicit approval. It suggests and scaffolds changes, but you have final control.
Which LLM does the assistant use? The assistant uses OpenAI GPT-4o and cannot be changed to other models.
Do deterministic actions still consume credits? Yes, all actions consume credits whether fired deterministically or based on LLM reasoning.
What features should be explored in addition to the new Builder and Script? Several powerful features to leverage:
  • Global Instructions Topic: Create universal guidelines (tone, escalation rules) once instead of repeating
  • Fallback Topic: Handle unmatched requests with knowledge search to reduce hallucinations
  • Custom Search Index: Create custom search index and retriever instead of using auto-generated defaults
  • Testing Center Integration: Systematic testing against prepared question sets builds customer confidence

 

Vidensartikelnummer

005232662

 
Indlæser
Salesforce Help | Article