Companies have a natural urge to get things under control, and standardize. Resist that urge for AI coding agents.

It’s a huge pain that everyone is using different coding agents. Cursor, Codex, Claude Code, Amp, etc, use different rules files, modes, prompting techniques, and sandboxing models.

And yet: it’s too early to pick one for the whole company.

You think that, through standardization, everyone can work in the same direction in lockstep. Less wasted effort; less thrash.

But things are moving too fast. For now, the benefits of staying loose and exploratory outweigh the benefits of standardization. Standardizing requires that you pick one gold standard — say, Claude Code — and have everyone adopt it.

Sure, everyone is going to have marginal efficiency gains from knowing to use CLAUDE.md instead of AGENTS.md. But my personal preferences and learnings around coding agents change every week, if not every day. A few weeks ago, I loved Cursor. But then, it was replaced by Claude Code on the web. Or what about Amp?

If you mandate (or “strongly recommend”) a singular coding agent to your engineers, you won’t get the breadth of exploration that will let you spot the next great tool. Claude Code is great today, but that doesn’t mean that something meaningfully better isn’t around the corner.

So instead of picking winners, let interested employees explore new tools. If any of those new tools represent a meaningful improvement, you won’t want to miss it for the sake of unified billing or one less rules file.