Design the agent loop, not the prompt (harness engineering)

ExplainX·medium signal

Anthropic engineers (Boris Cherny) are pushing 'harness engineering': stop hand-writing prompts and instead design the loop the agent runs—gather context, act, verify its own work, repeat—plus the tools and guardrails wrapped around it. The practical shift is to invest in the verification step and the stopping condition rather than the wording of any single prompt; a good loop with a mediocre prompt beats the reverse. Treat it as a first-class skill alongside prompt and context engineering.