Native function calling in LLMs and the use of agents serve distinct yet complementary roles in the deployment of AI systems. Function calling is best for direct, single-step actions linked to external functions, while agents are suited for more complex, interactive, and multi-step scenarios.

Definitions

Function calling in LLMs refers to the model’s ability to invoke functions from external APIs or libraries based on user prompts. Agents in LLMs are models designed to perform a series of tasks, often involving multiple steps and decision-making capabilities.

Use Native Function Calling Over Agents When:

• The task is straightforward and involves direct data retrieval or single-step commands
• There is a need for high precision and reliability in executing well-defined functions
• Integration with specific external systems or APIs is a priority, and the tasks do not require contextual understanding or multi-step logic

Use Agents (With or Without Function Calling) When:

• Tasks are complex and involve multiple steps or require planning and adaptation based on dynamic inputs
• Ongoing interaction and context management are crucial, such as in interactive customer service or complex problem-solving environments
• The application benefits from an autonomous system that can understand and navigate workflows or processes independently

How Do They Perform?

Function calling is a critical skill for LLMs, especially when integrating with external APIs or databases. However, LLMs can struggle with tasks that require real-time data or complex multi-step processes. Agents on the other hand are designed for multi-step tasks and decision-making, can handle complex workflows but may fail over longer trajectories or in tasks requiring extensive planning.

• A talk on the paradigm of function calling
• Andrew Ng’s talk on agentic workflows
• The original ChatDev paper (2023)