From Zero to One: Building AI Agents From the Ground Up | Federico Bianchi, TogetherAI

Item: From Zero to One: Building AI Agents From the Ground Up | Federico Bianchi, TogetherAI
Rating: 4
Author: PostSingularity Institute

by Federico Bianchi

Applied Sciences & Engineering AI — News & Applications UYComputer science UYQArtificial intelligence (AI)

🎙 Federico Bianchi 👥 5K 📅 October 24, 2025 ⏱ 77 min 👁 111 📄 expert opinion 🧭 2026-08-15

Available in: English (current) Français

Keywords

agent architecturetool usesupervised fine-tuningreinforcement learningcode interpreter

Summary

Federico Bianchi, Senior ML Scientist at Together AI, presents a comprehensive overview of building AI agents from the ground up. He begins by deconstructing the term ‘agent’, noting the lack of a universally accepted definition and the importance of understanding core components. He outlines the fundamental agent loop: user input, LLM processing, tool calls, execution, and feedback. He distinguishes between agents (autonomous decision-making) and workflows (predefined paths). The talk then focuses on three methods for building agents: ReAct (reasoning and acting) with structured prompting, CodeAct (using Python code as actions) for more flexible tool use, and supervised fine-tuning (SFT) to learn new behaviors from distilled examples. Finally, he discusses reinforcement learning (RL) to further improve agent behaviors through reward-based learning. Throughout, he emphasizes the importance of safe code execution environments, such as Together Code Interpreter, and provides practical insights from his experience. The talk concludes with a brief mention of an AI-organized conference, highlighting the potential of agents in scientific discovery.

161 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable practical insights into building AI agents, drawing from the speaker’s hands-on experience. The argumentation is clear and logical, progressing from basic concepts to more advanced training techniques. The speaker effectively explains the trade-offs between different approaches, such as using Python code versus standard tool calls, and emphasizes the importance of safe execution environments. The inclusion of a live demo and concrete examples strengthens the practical value. However, the talk is more of an overview than a deep dive, and some claims lack rigorous empirical evidence.

98 words

Title / Content Match

The title accurately reflects the content: a comprehensive guide to building AI agents from scratch, covering architecture, frameworks, and training methods.

Quality & Reliability

8/10

The speaker is a senior ML scientist at Together AI with relevant experience and publications. The talk provides practical insights and references to established frameworks (ReAct, CodeAct) and tools (Together Code Interpreter). However, it is a conference talk with limited depth and no formal citations or peer-reviewed validation of the presented methods.

Key Moments

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

Cited Sources

MLOps World — Conference website where the talk was recorded

Concurring Sources

ReAct: Synergizing Reasoning and Acting in Language Models — The paper that introduced the ReAct framework, which the speaker discusses.
Executable Code Actions Elicit Better LLM Agents — The paper that introduced CodeAct, which the speaker discusses.

Contribution & Novelties

The talk provides a practical, hands-on perspective on building AI agents, synthesizing common patterns and offering actionable advice. It highlights the importance of safe execution environments and the trade-offs between different agent architectures. The speaker also introduces an innovative AI-organized conference, showcasing the potential of agents in scientific discovery.

Pour aller plus loin :

ReAct: Synergizing Reasoning and Acting in Language Models — The foundational paper for the ReAct framework.
Executable Code Actions Elicit Better LLM Agents — The CodeAct paper, which introduces using Python code as actions.
Together Code Interpreter — Official documentation for the tool mentioned in the talk.

101 words

Radar Profile

The radar profile shows high scores in information quantity and quality, reflecting the speaker's expertise and the comprehensive coverage of the topic. The technical level is moderately high, suitable for an audience with some background in AI. The reliability score is slightly lower due to the lack of formal citations and the anecdotal nature of some claims.

Reliability 7/10