Choose the right AI model for each part of your agent to balance response speed, accuracy, and reasoning depth.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Performance, Unlimited, and Developer Editions. Required add-on licenses vary by agent type. |
| User Permissions Needed | |
|---|---|
| Manage AI Agents AND the required permissions for your agent type | |
How Model Override Works
Use model override when a subagent's specific job benefits from a different balance of speed and reasoning depth.
When a subagent runs, Agentforce determines which model to use in this order:
- The model set on the subagent
- The model set for the agent
- The model selected in Setup. By default, every agent uses the model specified on the Einstein Audit, Analytics, and Monitoring Setup page. To change the default model, see Select Agentforce Model Option.
For example, let's say your default model in Setup is GPT-4.1, your agent's model isn't set, and you select Google Gemini 2.5 Flash on the General FAQ subagent. Agentforce uses Google Gemini 2.5 Flash when the General FAQ subagent runs, but it uses GPT-4.1 for all other subagents.
Model override doesn't affect the model that prompt templates in actions use.
Choosing a Model for a Subagent
The model that a subagent uses affects how reliably it follows instructions, how fast it responds, and how well it handles complex tasks. When choosing a model, consider these factors:
- The model's context window determines how much the model can process in one session. A model with a small context window can lose track of earlier conversation turns or drop instructions in long interactions.
- Smaller models are faster but are more likely to generate inaccurate information, misinterpret complex instructions, and produce malformed structured output. They also tend to ask more clarifying questions or add small talk.
- If your subagent only processes text, a text-only model is often faster than a multimodal model.
| Subagent's job | Model guidance |
|---|---|
| Simple routing or classification | Use the HyperClassifier model, a Salesforce model optimized for subagent classification, or a similar small model. For more information, see Subagent Classification and Routing. |
| Looking up records or performing simple tasks | Use a small model that reliably follows instructions and calls the right action. |
| Complex reasoning, multi-step analysis, or long-form generation | Use a large model with strong reasoning capabilities. |
| Structured output | Use a large model that handles structured output reliably. |
| Image or document processing | Use a model with multimodal support. For a list of available models, see Large Language Model Multimodal Support. |
For a list of available models and their capabilities, see Supported Models. The Einstein Trust Layer applies to all model calls.
Test how the subagent performs in the Preview panel and Testing Center.
- Run the same conversations with different models and compare response quality, directness, and speed.
- Start with a capable model to set a baseline. Then try smaller models to find the smallest one that meets your business needs.
- Test edge cases. Models handle vague or incomplete messages differently.
Considerations
- Model override is available only for agents created in the new Agentforce Builder.
- For information about model deprecation, see Prepare for Model Deprecation and Rerouting