A harness for seniors and juniors

I came across a paper on how the harness, meaning the capabilities available in an agent (Claude Code, Codex, and so on), affects LLM performance.

The results are telling: a heavy harness hurts strong LLMs or is redundant, while weaker ones absolutely need a complex harness.

An interesting parallel: a senior developer can get by with a command line, vim, and grep, while a junior can't imagine working without a GUI and an IDE.