Uploaded March 2026 | Updated September 2026, 3 weeks ago
Someone improved 15 different LLMs at coding — all of them — in a single afternoon. They didn't fine-tune anything. They didn't switch models. They changed everything around the model.
That's harness engineering — the four layers between your intent and the model's output that actually determine your results.
00:00 Intro
00:14 How We Got Here
01:08 What Is Harness Engineering
01:56 My Personal Setup
02:36 Layer 1: Agent Frameworks
03:43 Layer 2: Tool Harnesses
04:45 Layer 3: Prompt Harnesses
05:55 Layer 4: Inference Engines
06:52 The Convergence
08:15 What To Do Next
🧠 Layer 1: Agent Frameworks
How your model plans, acts, and recovers. Comparing OpenAI Agents SDK, Google ADK, Anthropic Agent SDK, LangGraph, CrewAI, and AutoGen.
🔌 Layer 2: Tool Harnesses
MCP (Model Context Protocol) — the USB-C of AI. 97M monthly SDK downloads, 10K+ community servers, adopted by every major player.
📄 Layer 3: Prompt Harnesses
Beyond system prompts — hierarchical instruction files, DSPy auto-optimization, Microsoft Guidance constrained generation.
⚡ Layer 4: Inference Harnesses
vLLM vs SGLang benchmarks, PagedAttention, Ollama, llama.cpp — where and how fast your model runs.
The models are converging. The harnesses are diverging. That's where builders win or lose.
📺 More deep dives coming on MCP and agent frameworks — subscribe to catch them.
Someone improved 15 different LLMs at coding — all of them — in a single afternoon. They didn't fine-tune anything. They didn't switch models. They changed everything around the model.
That's harness engineering — the four layers between your intent and the model's output that actually determine your results.
00:00 Intro
00:14 How We Got Here
01:08 What Is Harness Engineering
01:56 My Personal Setup
02:36 Layer 1: Agent Frameworks
03:43 Layer 2: Tool Harnesses
04:45 Layer 3: Prompt Harnesses
05:55 Layer 4: Inference Engines
06:52 The Convergence
08:15 What To Do Next
🧠 Layer 1: Agent Frameworks
How your model plans, acts, and recovers. Comparing OpenAI Agents SDK, Google ADK, Anthropic Agent SDK, LangGraph, CrewAI, and AutoGen.
🔌 Layer 2: Tool Harnesses
MCP (Model Context Protocol) — the USB-C of AI. 97M monthly SDK downloads, 10K+ community servers, adopted by every major player.
📄 Layer 3: Prompt Harnesses
Beyond system prompts — hierarchical instruction files, DSPy auto-optimization, Microsoft Guidance constrained generation.
⚡ Layer 4: Inference Harnesses
vLLM vs SGLang benchmarks, PagedAttention, Ollama, llama.cpp — where and how fast your model runs.
The models are converging. The harnesses are diverging. That's where builders win or lose.
📺 More deep dives coming on MCP and agent frameworks — subscribe to catch them.
