Harness Engineering 101
The LLM is the brain. The harness is the body.
Everyone talks about AI agents like they’re a new kind of software: tool calls, memory, planning, RAG, MCP, multi-agent orchestration. Stack enough of those words together and it starts to feel like there’s some hard machine humming underneath that you’re supposed to already understand.
There isn’t. Here’s the whole thing:
An LLM API is stateless. Every turn, you send the entire conversation as a JSON array and get text back. A harness is the program that builds, maintains, and protects that array. Everything the field calls “agents” is a set of patches to that one loop, and each patch exists because something concrete broke.
That’s the claim this series defends. Every “agent” feature you’ve heard of is one patch to that one loop, and each patch shows up because something concrete broke. Learn the patches in the order they were invented, as fixes to real failures, and none of it is complicated. No frameworks, no buzzword tour, no architecture diagrams with twelve boxes.
I’m not guessing at this. I started building generative AI systems when GPT-3.5 was the frontier: no tool calls, no caching, no agents. You used the OpenAI SDK to send a message array and got formatted text back. Everything added since was added for a reason, and I watched most of it show up one patch at a time.
So that’s the shape of every chapter: here is the failure, here is the minimal patch, here is what the patch costs you. A toy harness in plain Python (raw HTTP, no SDK) grows alongside the text. By the epilogue you will have a working mini coding agent in about 300 lines, and the durable knowledge that there is no magic anywhere in the stack. The full source is on GitHub.
Two real harnesses show up along the way as proof the patterns hold: Claude Code, and One Code, my reimplementation of it on another agent runtime. They are here to confirm the patterns work in production, not to do the explaining. The toy harness does that.
Read the series
The full series is a short book: sidebar, search, and all sixteen chapters plus the appendices in reading order. It grows a toy harness in plain Python (raw HTTP, no SDK) alongside the text; by the epilogue it’s a working mini coding agent in about 300 lines.
Read Harness Engineering 101 →
The runnable harness and every chapter’s source are on GitHub.