What if we've been making our AI coding agents worse by giving them more rules? 🤯

An ETH Zurich study challenges the value of AGENTS.md files.

The findings:

📉 No overall improvement in task success
📈 Over 20% higher inference costs
🤖 More instructions don't automatically mean better results.

This really got me thinking.

I use AGENTS.md myself while developing my Unity game, Drone Harvest, with Codex.

Until now, I've always assumed that more detailed instructions and documentation would make my agents more reliable.

But what if I'm actually getting in their way?

Maybe the future isn't about writing longer instruction files, but keeping only the rules that truly matter for each project.

What about you? Are your AGENTS.md files getting longer, or have you already started trimming them down? 👀

Study: https://arxiv.org/abs/2602.11988

#Codex #ClaudeCode #AIAgents