Two papers published in early 2026 suggest you might have just made your agent slower, more expensive, and no more accurate. The right mental model is to tre...

Addy's platform is a resource for web developers and performance enthusiasts, offering insights into web performance optimization, browser rendering, and front-end development techniques. Through articles, tools, and case studies, Addy offers insights into measuring and improving web page speed and user experience. Developers can learn about performance budgets, lazy loading, and optimizing web assets to create fast and responsive websites.

Addy Osmani

Auto-generated AGENTS.md files (produced via /init) hurt AI coding agent performance and inflate costs by 20%+ because they duplicate information agents can already discover by reading the codebase. Two 2026 research papers show LLM-generated context files reduce task success while increasing cost, whereas human-written files help only when they contain non-discoverable information like tooling gotchas, non-obvious conventions, and hidden landmines. The right mental model is to treat AGENTS.md as a minimal, living list of codebase friction points that can't be inferred—not a comprehensive onboarding document. Every discoverable line is noise that competes with the actual task via context dilution. A better architecture involves a routing layer with dynamically loaded, task-specific context rather than a monolithic static file, though tooling support for this is still lacking.

Stop Using /init for AGENTS.md

The better architecture people aren’t building yet

I highly question this. I was running pi-coding-agent from a directory that houses several repos today. One in particular, is a legacy code base. The others are the modern implementations. The reason for running it from that directory was that we’re still porting the last few features and, not by choice, one of them is incredibly complex.
Without the agent file, the AI was struggling to figure out anything at all. But even once I defined directory names and structures, it still took several iterations of “polishing” before the experience became anything close to smooth.
Should I have had the agent write the file? Maybe. But not having one at all made it impossible to work with. So this point that having an agent file that just points out gotchas and such… I don’t buy it.

the 20% cost inflation thing is lowkey insane when you think about it at scale. like imagine running hundreds of agent tasks daily and just burning tokens on context the model couldve figured out itself. honestly tho im curious where yall draw the line between “the agent cant infer this” vs “just trust the model bro” because that line feels super blurry in practice