Building AI Agents The Better Way
Most developers are building AI agents the wrong way. Creating limited agents with simple text input and text output. This approach limits the potential of what AI agents can accomplish.
Watch this 5-minute video that demonstrates the implementation.
The Better Approach
The right way to build AI agents leverages function/tool calling for deterministic JSON output format. This is a critical aspect to creating AI coding agents or any other kind of AI agent, and it enables you to automate virtually any workflow.
Building AI agents using this approach with function calling is like giving the AI a fixed list of possible responses instead of just letting it talk. The old way of building AI agents meant the AI could only respond with words and explanations, but couldn't actually do anything. It couldn’t perform tasks. It couldn’t classify intent using your own specific schema. The AI that describes what a cat looks like but can't draw one.
The batter approach uses function calling. This gives the AI agents the ability to perform real actions, return information in organized formats like JSON, complete multiple tasks in a row without constant guidance, and work together with other systems and tools.
This approach enables real applications like the Waldo Finder (below) that can locate and mark objects in images. The Chart Reader (below) that can analyze graphs and metrics. And the Auditor (below) that can examine data to find mistakes and inconsistencies, can even be enabled to make corrections without a human.
Instead of building a chatbot that only talks about solutions, developers can now build AI agents that actually solve problems and get work done.
Why Function Calling Matters
Function calling provides:
Deterministic output format - Structured, reliable responses in JSON
Workflow automation - Ability to chain actions and automate complex processes
Structured data handling - Direct integration with APIs, databases, and other systems
Platform Support
This approach is well-supported across major AI platforms:
Anthropic Claude
Google Gemini
OpenAI GPT
gpt-oss models
Real Examples
Here are AI Agents built using this approach:
Waldo Finder (Claude)
An AI agent that can locate and identify specific elements or patterns.
Chart Reader (Claude)
An AI agent that reads Dashboard Usage Charts and has integrated knowledge from internal documentation, enabling it to provide informed analysis of metrics and usage patterns.
AI Auditor Agent (OpenAI)
The following AI Agent can read JSON data files and find data corruptions and pollution in the data. The AI responds with exact record IDs and reasoning:
