Six New Tips for Better Coding With Agents

Half of a baker's dozen: Steve and Gene ran through thirteen-odd new observations one morning on the road, and this post ships only the ones that "seemed the most baked" — the rest, Steve admits, don't have vocabulary yet. The hurry shows, and it suits the material, since the practice was outrunning print. The opening act of heresy could carry the piece alone: Joel Spolsky's 1999 commandment against ever rewriting working software, which Steve rates among the best advice anyone has given a programmer, gets declared expired after twenty-six years of faithful service. The rest is craftier and more specific: why agents refuse the tools people lovingly build for them, a borrowed "Rule of Five" for reviews that sounds absurd right up until it works, and hard numbers on how much of a coding budget code health should eat.

What keeps the advice honest is the failure count. For every trick that works, Steve shows one that didn't: the agent-Emacs bridge that Claude designed and then refused to touch ("Claude never uses our fuckin' bridge we built together"), the flawless swarm run he could not reproduce afterward, the orchestrator still hunting for its magic formula. The swarming section pivots on a small mystery — thirty finished, carefully groomed issues vanish from his tracker — that this note won't solve. A year on, readers can also score the predictions: an orchestrator promised for Christmas Day, and a 2026 populated by engineers running a hundred agents apiece.