Why Salesforce Introduced the New Agent Builder
The legacy builder relied almost entirely on prompts, which made agents unpredictable and gave admins little visibility into their behavior.
The new builder uses hybrid reasoning: you decide which steps the LLM reasons through and which follow fixed business rules.
Advantages Over the Legacy Builder
New Agent Builder Options
Agentforce Studio provides different tools to build, test, monitor, and manage AI agents.
| Option | What It Is | When to Use |
|---|---|---|
| Agents | Create and manage Agentforce Agents. | Create, configure, test, activate, and manage agents. |
| Tests | Central place to test agent behaviour. | Check responses, Subagent selection, Actions, and expected results. |
| Grids | Grid-based workspace to process multiple records. | Use AI with many records in a structured grid format. |
| Prompt Templates | Opens Prompt Builder. | Create Record Summary, Field Generation, and Flex prompts. |
| Data | Manages data sources used by Agentforce. | Knowledge, Data Libraries, Search Indexes, Retrievers, and grounding. |
| AI Models | Manages AI models available for Agentforce. | Configure or select supported AI models. |
| Agentforce DX | Developer-focused development tools. | VS Code, CLI, source-driven development, and deployment. |
| Analytics | Agent performance and usage information. | Understand usage, performance, outcomes, and trends. |
| Sessions & Intents | Agent conversations, sessions, and user intents. | Troubleshoot conversations and understand user requests. |
| Alerts | Important Agentforce warnings and issues. | Identify failures, unusual behaviour, and issues. |
Exploring the New Agent Canvas
4.1 Top Bar
| Option | Purpose | When to Use |
|---|---|---|
| Agent Name + Version | Displays the agent and version being edited. Draft = still editable. | Switch between versions. |
| Save | Saves draft changes. | While editing the agent. |
| Commit Version | Locks the draft as a fixed version. Create a new draft for further edits. | After testing is complete. |
4.2 Explorer — Left Panel
- Search: Find elements such as Subagents, Actions, variables, or connections.
- Agent Definition: View or edit the complete Agent Script blueprint.
Explorer Settings
| Setting | What It Holds | When to Use |
|---|---|---|
| Agent Settings | Name, API name, description, running user, general setup. | First, while creating the agent. |
| System Messages | Welcome message and error message. | Control greeting and fallback wording. |
| Agent-Level Instructions | Global rules for all Subagents. | Common tone, persona, formatting, and restrictions. |
| Language Settings | Primary and supported conversation languages. | When users chat in multiple languages. |
| Variables | Record ID, user information, flags, and stored values. | Pass data between Subagents and Actions or evaluate conditions. |
| Connections | Deployment channels and surfaces. | Before making the agent available to users. |
| Data | Data Libraries, Knowledge articles, and files. | When answers must use trusted information. |
Rule applies everywhere → Agent-Level Instructions
Rule applies to one job only → Subagent Instructions
4.3 Preview Panel — Center
- Preview / Agent Def: Switch between agent chat and script.
- Info Icon: Explains preview modes.
- Refresh Icon: Resets the conversation and starts a new session.
- Chat Box: Test messages, off-topic questions, and unclear requests.
Preview Modes
| Mode | What It Does | When to Use |
|---|---|---|
| Simulate Mode | AI pretends to execute Actions. No real data is read or changed. | Early testing of routing and logic. |
| Live Test Mode | Executes real Actions against your org. | Final checks using real records and functionality. |
4.4 Agent Preview Details
- Context Variables: Set values before chatting, such as the current user or Record ID.
- Interaction Summary: Review which Subagent was selected, which Actions ran, and the reasoning.
4.5 Bottom Bar
| Option | Purpose |
|---|---|
| Trace | Step-by-step execution details for debugging Subagent selection and Action inputs/outputs. |
| Variables | Live values of variables during the conversation. |
| Session ID | Unique test session identifier for logs and issue reporting. |
| Expand / Collapse | Show or hide the trace panel. |
4.6 Agentforce Panel — Right
This is an AI assistant that helps you build the agent. It is not the agent being tested.
"Add a subagent for case escalation"
"Why did my agent choose General FAQ?"
"Explain this trace step"
The Refresh icon resets the assistant; X closes the panel for more screen space.
Modes of the New Agent Builder
Open the new Agentforce Builder from App Launcher → Agentforce Studio.
The old builder, accessed through Setup → Agentforce Agents, is known as the Legacy Builder.
Chat View
Build using plain English.
- Chat with Agentforce.
- Describe the required functionality.
- Generate Subagents, Actions, and logic.
- Explain problems and request changes.
Canvas View
Low-code agent development.
- Default builder view.
- Document-style editing.
- Easy-to-read script blocks.
- Shortcuts for Actions and conditions.
Script View
Pro-code agent development.
- Edit actual Agent Script.
- Syntax highlighting.
- Auto-complete and validation.
- VS Code / Agentforce DX support.
All three approaches work with the same Agent Script definition. Changes in Canvas and Script stay in sync.
What Is Agent Script?
Agent Script is a language that combines natural-language instructions for an LLM with fixed logic for business rules.
What Agent Script Supports
- If/else conditions
- Setting and updating variables
- Moving between Subagents
- Calling Actions
- Controlling business execution
| Syntax | Purpose |
|---|---|
-> |
Logic part |
| |
Prompt / instruction part for the LLM |
Can We Build Agents with Only Canvas?
Canvas supports Subagents, instructions, Actions, variables, conditions, and transitions.
Canvas also supports deterministic logic through shortcuts and can be used together with the Chat assistant.
When Script View Is Helpful
- Complex logic with nested conditions.
- Action chaining in an exact order.
- Fine control over variables and transitions.
- "Available when" filters.
- Reusable patterns and recipes.
- Debugging detailed agent behaviour.
- Source control / CI-CD with VS Code.
- Advanced MCP Action developer names.
Use Canvas to build the structure, then use Script for precise logic and fine-tuning.
Subagents and Actions — Limitations
From April 2026, the term Topics is called Subagents.
Count Limits — Commonly Quoted
| Component | Approximate Limit |
|---|---|
| Active Agents per Org | About 20 |
| Subagents per Agent | About 10 |
| Actions per Subagent | About 10 |
Time Limitations
| Component | Limitation | Impact / Solution |
|---|---|---|
| Agent Action | Around 60 seconds | Long-running Actions may fail. Keep execution fast or use asynchronous processing. |
| External Systems | May be slow | Slow external or legacy systems can exceed Action timeouts. |
| Apex / Flow | Normal governor limits apply | SOQL, DML, CPU, and other platform limits must be considered. |
Best Practices
- Use specialised agents and an Orchestrator Agent for larger solutions.
- Keep Actions efficient.
- Use asynchronous processing where appropriate.
- Consider MuleSoft / MCP for cross-system work.
- Bulkify and optimise Apex.
- Avoid unnecessary repeated Action calls.
Evolution of Salesforce Agents
Einstein Bots
Rule-based chatbots, fixed dialogs, no real reasoning.
Einstein GPT
Generative AI in Salesforce for emails and summaries.
Einstein Copilot
AI assistant for Salesforce users with Topics, Actions, and reasoning.
Agentforce 1.0
Autonomous agents, Atlas Reasoning Engine, and Agent Builder (now Legacy Builder).
Agentforce 2.0
More prebuilt skills, Slack support, better reasoning, and data retrieval.
Agentforce 3.0
Command Center, observability, MCP support, and more partner Actions.
Agentforce 360
New Agentforce Builder beta, Chat / Canvas / Script views, and Agent Script with hybrid reasoning.
New Agent Builder
Subagents terminology, connected Subagents, multi-agent design, and Agentforce DX support.