Build-Your-Own Coding Harness: The 200-Line Core Loop
A minimal AI coding harness is a 60-200 line program that wraps the tool-call loop around three tools — read_file, list_files, edit_file — or even just bash alone, because modern models are trained on tool-call transcripts.
Modern AI coding harnesses are smaller than they look. Mihail Eric's writeup "The Emperor Has No Clothes: How to Code Claude Code in 200 Lines of Code" walks through a minimal coding harness whose core loop is 60-200 lines of Python. The loop is just: send prompt + system prompt listing tools, parse the model's tool call, run it, append the tool result to the transcript, re-invoke. See Tool Calling Loop: How a Coding Harness Drives a Stateless Model for the underlying mechanic. Three tools are sufficient to build a useful agent: read_file, list_files, and edit_file. Even more starkly, bash alone is enough — modern models have been trained on enough tool-call transcripts that they will discover the right shell incantations to read, list, and edit on their own. A live-demo strip-down by Theo Browne (t3.gg) compressed a working harness to about 75 lines. A related artifact in this space is T3 Code, launched March 6, 2026 as open source (pingdotgg/t3code), BYOK. T3 Code is not a harness in the sense above — it ships zero tools and is a UI wrapper that requires Claude Code CLI or Codex CLI to be installed locally. The harness is still the underlying CLI; T3 Code just provides chrome. The practical upshot is that harness lock-in is mostly about subscription billing rather than technical complexity. Anthropic and Google tie subscriptions to their own harnesses; OpenAI lets subscribers use credits across third-party harnesses. So if the Matt Mayer Harness Benchmark: Same Opus, 16-Point Swing Between Claude Code and Cursor motivates you to run Opus inside Cursor, you have to pay Cursor separately rather than reuse your Claude Max subscription.