In OpenAI's experiment the standing instructions were kept deliberately tiny. Everything else that the tool needed to know was written down separately, and kept where the work was. Note the scale below: each step along the axis is ten times the one before it.
3 to 7
People on the team. Output per person rose as it grew, rather than falling.
1,500
Separate changes made and accepted over five months.
1 day in 5
Spent tidying up by hand, until the standards were built into automatic checks instead.
Source: Ryan Lopopolo, "Harness engineering: leveraging Codex in an agent-first world", OpenAI, February 2026. Figures as reported. Line counts are approximate and stated by the author as being of that order.