How I Built an AI Agent with LLM Function Calling (and Avoided Unnecessary Tool Calls)

Description and use

Built an AI agent using LLM function calling with a decision layer to avoid unnecessary backend calls, reducing latency and improving efficiency.

Problem

The LLM was calling backend APIs on almost every user input, causing increased latency, unnecessary API calls, and inefficient system behavior.

Outcome

Reduced unnecessary tool calls, lowered latency, and improved response consistency in the AI-powered course recommendation system.

Workflow

• 1. Receive user input
• 2. LLM generates intent and decision to call a tool
• 3. Orchestration layer evaluates decision
• 4. If tool call is needed, LLM gathers missing parameters via follow‑up questions
• 5. Execute backend API (tool) only when all parameters are present
• 6. LLM generates final response to user

Evidence

The model was calling backend APIs almost every time—even when it wasn’t necessary.

Source

Instead of blindly executing tool calls, the agent identifies missing information, asks follow‑up questions, and executes the function only when all parameters are available.

Source