Offloading Human Tasks with DeepMind's Maarten Grootendorst

Join us at AGNTCon + MCPCon North America, October 22–23, 2026 in San Jose: https://events.linuxfoundation.org/agntcon-mcpcon-north-america/ If you build AI agents, much of the work you do by hand today is getting trained directly into the models: step-by-step reasoning prompts, tool-call formatting, and parsing agent loops. Maarten Grootendorst, co-author of the best-selling O'Reilly book Hands-On Large Language Models and the new An Illustrated Guide to AI Agents (with Jay Alammar), shows how each core piece of an agent moved from prompt engineering to trained behavior, and what that means for the harnesses we build next. In this keynote from AGNTCon + MCPCon Europe 2026 in Amsterdam, Maarten covers: - From chat to action: How the field moved from LLMs that talk to LLMs that take actions in just a couple of years - The four parts of an agent: A reasoning LLM augmented with memory, tools, and planning - Reasoning, trained in: Why "let's think step by step" gave way to dedicated reasoning and answer tokens - Native tool calling: From hand-written XML or JSON prompts and abusing the user role to models trained to call tools - The ReAct loop: Reason, act, observe, and why running it by prompt alone meant a lot of manual parsing - Agents by training data: Add a little ReAct data during training and you get an agent out of the box - What this means for harnesses: Predicting which parts of today's agent code future models will absorb - What's still done by hand: Swarms, guardrails, and evaluations as the next things to be trained in For AI engineers and agent builders who want to know which parts of their stack are about to become the model's job. Links & Resources: - An Illustrated Guide to AI Agents (Maarten Grootendorst & Jay Alammar): https://maartengrootendorst.com/book/ - An Illustrated Guide to AI Agents on O'Reilly: https://www.oreilly.com/library/view/-/9798341662681 - Hands-On Large Language Models (O'Reilly) - AGNTCon + MCPCon North America: https://events.linuxfoundation.org/agntcon-mcpcon-north-america/ - AGNTCon + MCPCon Europe: https://events.linuxfoundation.org/agntcon-mcpcon-europe/ - Agentic AI Foundation: https://aaif.io Timestamps: 00:00 From LLMs that chat to LLMs that act 00:38 From Hands-On LLMs to An Illustrated Guide to AI Agents 01:41 The four parts of an agent: reasoning, memory, tools, planning 02:00 Reasoning: from prompting to trained reasoning tokens 03:00 Tool calling: from JSON prompts to native tool use 03:57 Planning and the ReAct loop 05:32 Why prompted ReAct needed so much manual parsing 05:57 Sprinkle in ReAct data and you get an agent 06:25 What this means for agentic harnesses 07:24 Next frontier: swarms, guardrails, evaluations 07:58 We are just starting the era of agentic AI #AIAgents #ToolCalling #ReasoningModels