Harness Engineering

Harness Engineering Table of Contents Introduction A simple harness The core loop The execution log Build the context Add tool call support LLM Support Observability Big picture What's next Takeaways In the first couple of years of LLMs, we talked a lot about "prompt engineering": how to craft a prompt just so in order to induce the LLM to do what we want. Since then, we've moved on to the more expansive "context engineering": how multiple available sources of context (chat history, file reads, web search results, etc.) should be combined to induce an agent to do what we want. Here, I'm writing about the even-more-expansive notion of "harness engineering": how do we orchestrate the whole loop of LLM invocations, tool calls, background process management, context retrieval, permission asks, etc., to get the agent to do what we want? Introduction I have been using Claude Code and OpenAI Codex for months now. The underlying models clearly drive much of the effectiveness of these…