One engineer ran 50 to 60 AI coding agents on a single codebase for ten weeks. The agents did not end up with a better sandbox. They ended up with a body of law. This is one engineer's account of his own system, not a measurement.
Steve Yegge, a veteran engineer and long time writer on software, published the account on August 24, 2026. His agents averaged 270 commits a day and, left alone overnight, made at least one bad decision every night. What they built to cope was 450 written rules, more than 100 of them fences. A fence, in his definition, is any mechanism that turns you away if you are not supposed to be there: a polite refusal, not a wall. Rules follow a lifecycle: a custom, then a warning when broken, then written law every agent must obey, then mechanical enforcement, a program that refuses by policy or alerts loudly. The results he reports: a game that used to go down for days at a time now has alerts on everything, its production infrastructure rewritten, and a relaunch on three platforms ready in under ten weeks.
Inside a client's team we see the same lifecycle at a smaller scale. A reviewer catches something twice. It becomes a line in the context file. When it is caught a third time it becomes a check that runs before the pull request opens, that is, before anyone reviews the code.
Harvard Business Review on X: adoption improves when organizations are explicit about what an agent may and may not do.
Sources
- Steve Yegge, Fences, not Sandboxes, August 24, 2026. https://yegge.ai/essays/fences-not-sandboxes/
- Harvard Business Review on X, September 10, 2026. https://x.com/HarvardBiz/status/2097884015944442217