Nir Kaufman on Building AI Agents with TypeScript, Tools and Human Control

About this episode

A chatbot can give you an answer. An AI agent can take action. What changes when we give software the tools to do things for us?In this episode of AI Master Podcast, I sit down with Nir Kaufman to talk about building agents with TypeScript and keeping people in control.We discuss the agent loop, clear tool descriptions, choosing models for different tasks, structured output, memory and context. We also explore RAG, middleware, MCP, and the design choices behind single-agent and multi-agent systems.One of my favourite examples involves a pizza. When should an agent ask a person for approval? Ideally, before the pizza goes into the oven. It is a simple way to think about actions that are difficult to undo.We also talk about evaluation and tracing, the mindset behind Nir's hands-on workshop, and why music, hiking and real human connections still matter in a world full of AI.Nir brings more than 20 years of software development experience and a real love of teaching and community. It was a pleasure to have him back for this conversation.Enjoy the episode, and let me know: what would you trust an AI agent to do on its own?Nir's paid workshop, Agent Development in Practice with TypeScript: A Deep Dive:https://ai-poland.pl/workshop/agent-development-in-practice-with-typescriptNir's free workshop, From Zero to Production Agent in 60 Minutes:https://ai-poland.pl/free-workshops/zero-to-production-agentExplore AI Poland:https://ai-poland.pl/More podcasts and stories from me:https://kalbarczyk.dev/Host: Dariusz KalbarczykGuest: Nir KaufmanLanguage: English