Harness Dojo: Practice the Loop

Field notes

The planned workflow begins with a weak loop and visible failures, then has the user inspect every message, tool call, result, and grader verdict before submitting a revision. The site specifies thirty hidden tasks, an 85 percent pass threshold, and a pinned model/environment so the harness—not model selection—accounts for the difference. A sample trace includes Python source context, token and step counts, and a displayed run cost of $0.018.