When I first started using AGENTS.md, I thought it was only useful for AI.

Over time, I realized that this is not the whole story.

The main benefit of this file is not just that AI makes fewer mistakes. It helps preserve the working context of the project.

Repeating the Same Instructions

Before using AGENTS.md, almost every new conversation started the same way. I had to explain what the project was, which stack it used, which conventions existed, and what should not be changed.

Now most of that information lives in the repository.

Unexpected Changes

Without context, AI often starts improving the project in its own way.

For example, it may:

  • suggest new libraries
  • change the directory structure
  • add unnecessary layers of abstraction
  • break existing conventions

When those rules are described in AGENTS.md, these suggestions become much less frequent.

Example

Sometimes, instead of a small function, AI suggests introducing a separate abstraction layer.

Before:

load_config()

Suggested:

class ConfigProvider:
    ...

For a specific project, this may be unnecessary. If AGENTS.md says that the project prefers simple solutions, suggestions like this become less common.

Returning to Old Projects

The most unexpected effect was not related to AI.

After a few months, every project has to be remembered again. Why was this stack chosen? Which decisions were already made? What is considered normal in this repository?

The answers are already in AGENTS.md.

Conclusion

Over time, I stopped treating AGENTS.md only as instructions for AI.

For me, it became short project memory.

It helps AI. But it also helps me restore context faster.