AI Agents In-Depth – Function Calling, MCP and Tool Use Under the Hood

Item: AI Agents In-Depth – Function Calling, MCP and Tool Use Under the Hood
Rating: 4
Author: PostSingularity Institute

by Alan Smith

Formal & Physical Sciences AI — News & Applications UYComputer science UYQArtificial intelligence (AI)

🎙 Alan Smith 👥 227K 📅 August 5, 2026 ⏱ 60 min 👁 4K 📄 expert opinion 🧭 2026-08-13

Available in: English (current) Français

Keywords

function callingMCPtool useagentic AILLM

Summary

In this talk, Alan Smith provides a detailed technical overview of how AI agents use tools, focusing on function calling and the Model Context Protocol (MCP). He explains that large language models (LLMs) are token predictors and do not directly call tools; instead, they select tools and specify parameters, while the agent application executes the calls. He demonstrates this with a Postman example using OpenAI’s GPT-4o, showing the JSON protocol for function calls. He then discusses frameworks like LangChain, Semantic Kernel, and Agent Framework, highlighting their roles in abstraction and model switching. He presents a pizza ordering agent built with Agent Framework, illustrating how the model handles multiple tool calls and non-deterministic behavior. He also covers a vibe-coding agent that creates websites, and explains how retrieval-augmented generation (RAG) has evolved to use tool calling for more intelligent search. The talk emphasizes the importance of tool descriptions and query rewriting, and touches on multi-agent solutions. Overall, it provides a solid understanding of the mechanics behind agentic AI.

166 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the inner workings of AI agents, demystifying function calling and tool use. The speaker’s arguments are well-supported by live demonstrations, which effectively illustrate the concepts. He clearly explains the separation between the LLM’s role in selecting tools and the application’s role in executing them, and he highlights practical considerations such as non-determinism and the importance of tool descriptions. The progression from basic function calling to more complex agentic RAG is logical and enhances understanding.

Scientific Rigor, Source Quality, Title Accuracy

The speaker demonstrates strong technical rigor through hands-on demos and accurate explanations of protocols. However, he does not cite external sources or references, relying solely on his expertise and live examples. The title accurately reflects the content, as the talk indeed covers function calling, MCP, and tool use in depth. No comments were provided for analysis.

151 words

Title / Content Match

The title accurately reflects the content, which covers function calling, MCP, and tool use in depth.

Quality & Reliability

8/10

The speaker demonstrates deep technical knowledge through live demos and clear explanations of underlying mechanisms. Claims are consistent with current AI practices, though no external sources are cited.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

NDC Conferences — Conference organizer and host of the talk
NDC Copenhagen — Specific conference where the talk was recorded

Concurring Sources

OpenAI Function Calling Documentation — Supports the explanation of function calling in OpenAI models.
Model Context Protocol (MCP) — Provides background on MCP, which is mentioned in the talk.

Contribution & Novelties

The talk provides a clear, hands-on explanation of how function calling and tool use work under the hood, which is often treated as a black box. It bridges the gap between high-level agent frameworks and the underlying protocol, making it accessible to developers. The emphasis on non-determinism and practical considerations adds value.

Pour aller plus loin :

OpenAI Function Calling Documentation — Official documentation on function calling.
Model Context Protocol (MCP) — Official site for MCP, a protocol for tool integration.
LangChain — Framework for building LLM applications.
Semantic Kernel — Microsoft’s framework for AI orchestration.
Retrieval-Augmented Generation (RAG) — Original paper on RAG.

104 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-rounded and informative talk. The strengths are in the quantity and quality of information, as well as technical depth, making it highly valuable for developers interested in AI agents.

Reliability 8/10