How to Build Your Own AI Agent (Like PewDiePie)

PewDiePie is building his own AI assistant, Odysseus, and training a local model, Ajax, to power it. In this video I break down what Odysseus is, how an agent harness works from scratch, what it actually means to run a model locally with llama.cpp and Hugging Face, and how you can fine-tune a small model like Qwen 3.5 9B on agent traces so it calls tools reliably in your own harness. --- 🔗 *Links* - Odysseus (GitHub): github.com/odysseus-dev/odysseus - Claude system prompts: platform.claude.com/docs/en/release-notes/system-prompts - Hugging Face models: huggingface.co/models - Hugging Face hardware settings: huggingface.co/settings/hardware - Llama macOS app (llama.cpp): github.com/ggml-org/Llama-macOS - llama.cpp: github.com/ggml-org/llama.cpp - Qwen3.5-9B (Unsloth GGUF): huggingface.co/unsloth/Qwen3.5-9B-GGUF - Kolibri 1: huggingface.co/Aleph-Alpha/Kolibri-1 - Qwen3.8-27B: huggingface.co/Qwen/Qwen3.8-27B - Agent trace datasets: huggingface.co/datasets - Post-training agents playlist: youtube.com/playlist?list=PLo2EIpI_JMQvQZm-kVlz4wY1vWF0LBcf5 --- 🤓 *Topics Covered* - What PewDiePie's Odysseus is and how it works - How AI agent harnesses and tool calling work - Running and fine-tuning small local models --- ⏱️ *Timestamps* 00:00 - Introduction: PewDiePie’s AI Agent Journey 01:19 - What is Odysseus? (Agent Harness Overview & Setup) 02:00 - The Odysseus UI 02:44 - Built-in Tools, Memory & Hardware Recommendations 03:50 - Live Demo: Mistral Large 4 04:28 - How AI Agents Actually Work (The Agent Loop Explained) 05:35 - Tools, Functions & Execution Flows 06:23 - What is Context: System Prompts, Tools & Memory 08:08 - Conversation Transcripts & Reasoning Steps 10:32 - Running Local Models vs. Cloud API Models 11:53 - Why Run Locally? (Privacy, Latency & Offline Access) 12:48 - Finding Open-Source Models on Hugging Face 13:50 - Setting Up Local Inference with llama.cpp 14:46 - Matching Models to Your Hardware (Hugging Face Settings) 15:28 - Choosing the Right Quantization (Q4 vs. BF16) 16:52 - PewDiePie’s "Ajax": Fine-Tuning Qwen 3.5 9B 19:24 - Sourcing Agent Traces & The OpenAI Reasoning Token Issue 20:45 - Finding & Sharing Agent Datasets on Hugging Face Hub 21:38 - Hugging Face Post-Training Masterclass Tutorial 22:15 - Recap: Two Ways to Train Your Own Agent Model 23:32 - Final Thoughts & Next Steps 👋 *Connect with me* My website: alejandro-ao.com/ X