Link

I read this article when it first came out, but I feel compelled to write a note about it.

It's one of the most measured assessments of coding agents that I've seen and is written by someone who knows their shit (he wrote the pi coding agent, which for better or worse powers OpenClaw).

His point is that agents have limited context, and even with appropriate context, will still make little mistakes that can compound complexity very fast. Humans also make mistakes. But agents can accumulate them much faster. So fast that even the most diligent human reviewer, could never catch them all in a reasonable amount of time. And when those mistakes compound, and a human actually needs to understand and fix them, you're fucked. Or you tell your clanker to add a feature or something, but the codebase has become such a mess, that it can't even figure out how to do it. And you certainly don't understand the codebase well enough to direct it in a meaningful way.

His advice for using agents intelligently?

• tasks can be scoped so the agent doesn't need to understand the full system.
• the agent can evaluate it's own work.
• the output isn't mission critical.
• you just need a rubber duck.

Let the agent do the boring stuff that won't teach you anything. Or help you try out something that would take too long otherwise.

And the main suggestion? The title of the post. SLOW THE FUCK DOWN.

Give yourself time to think about what you're actually building and why. Give yourself an opportunity to say, fuck no, we don't need this. Set yourself limits on how much code you let the clanker generate per day, in line with your ability to actually review the code.

Still write code. It introduces friction that helps you learn.