Blueprint of a Frontier Coding Agent

In late 2024, I opened my computer and started assembling what I thought was a coding agent. I had an LLM — Claude, running inside Claude Desktop. I had Model Context Protocol. I had a stack of MCP servers I had handpicked and connected one by one: GitHub for version control and code access, Playwright for browser automation, Sequential Thinking to help the model work through multi-step problems, Context7 for documentation retrieval. I had markdown files such as architecture docs, blueprint files, handoff notes, that I maintained manually to give my agent something resembling memory across sessions. I called it a coding agent. What I did not know at the time was that I had built a very primitive version of something the industry had not yet named: an agent harness.