A coding agent is a loop. The harness is the work.

Owain Lewis builds a minimal Go coding agent around a familiar loop: send the model a request, execute any tool calls, return their results, and repeat until the model stops. His practical demo adds file edits, shell commands, model routing, events, and a task-to-PR workflow. Why it matters: The model is only one layer. The useful behavior comes from the harness around it: what tools exist, which actions need approval, how context is managed, whether tests and reviews happen, and where a human takes over.

Try this: Write down one agent workflow you use today. Separate the model call from the harness rules around it: tools, permissions, verification, review, and the final human approval point. Fix the weakest boundary before adding another capability.