Add reusable AGENTS.md config

[swissspidy]

Opening this for discussion among the team :)

To-do:

• Mention that README.md must not be manually changed as it's auto-generated.
• Follow https://docs.github.com/en/enterprise-cloud@latest/copilot/tutorials/coding-agent/get-the-best-results

[mrsdizzie]

Just a rant here, but might as well be written: Aside from any other feelings about these tools, it doesn't feel exiting to encourage their use even more given the current amount of what I would consider slop pull requests and the automated responses to any feedback on them.

I understand you can't control what people use and do and that the point of this file is to actually get better results from those tools -- but the effect these tools are having on the project for me is that I spend much less time looking at pull requests because many of them feel like a demotivating waste of time that clearly are not reviewed well / at all before hand. Adding a file that takes the current problem and asks a machine "but pretty please don't do that" does not help with that feeling : )

I guess a more practical suggestion / feedback would be to also include a more firm project stance on submitting slop pull requests and pull requests for features that haven't really been discussed or approved since these tools make generating them much easier. There should be some consequence for generating slop and then asking a real person to take time and review it, or else I can't really see that changing.

[schlessera]

Thanks for the honest and constructive feedback, @mrsdizzie !

I fully understand your sentiment and can relate to it. But I think it's important to point out that a distinction should be made here. What we do not want to encourage is slop pull requests and careless code. The question whether we should encourage AI-augmented coding practices or not is already past discussion as far as I can tell.

Please allow me to share my view on what I think is currently happening: We're currently in a transition that is not reversible and that, basically, establishes a new higher-order interface for humans to communicate with machines. Just as we went from machine code to assembler, from assembler to C, and so on, we are now at a stage where natural language will slowly establish itself as the main language to communicate with machines for software engineering purposes. Yes, we still need code that can be compiled to run on the CPUs, but we don't need the humans to directly craft that code anymore. It was never efficient to begin with to let the humans learn how computers communicate, instead of the other way around, and we have been working towards a better approach for a while. But the technology was not up to par until recently. While the tooling is currently hit or miss, it is not only constantly improving but has already conquered a large segment of the engineering world.

And I understand that not everyone wants to share that view. Therefore, I would want to put systems in place that make sure that we do have full control as humans over the process still, but that the sloppy pull requests that keep increasing will not completely tie us down and remove any chance of working on the fun bits for us. The agents.md file that Pascal has submitted in his pull request is one of these systems by making sure that we can control the quality of AI-augmented code. We can also provide instructions for AI tooling in that file, like discouraging the AI systems to produce pull requests that we will just close off as meaningless anyway. At the same time, I'd like to add systems that use those same AI tools to review pull requests automatically and label or archive meaningless pull requests right away and require a human to do some specific action (like a specific type of comment) to get them in our queue in the first place. In this way, we have a fighting chance against increasing PR spam and making sure that we can focus on the important parts.

In a similar vein, I'd like to automate some of the review process to again let us humans focus on the interesting and fun bits of the process. As an example, anything that can be easily checked by something like a linter should be automatically flagged for change requests as it just means that the pull request submitter didn't take the time to properly review their PR to begin with.

All this to say that I don't think AI will go away, and if we don't proactively add tooling to our toolkit to deal with it, it will take up more of our bandwidth and have us be stuck in meaningless work, not finding the time to work on the important stuff.

I like your suggestion with the firmer stance on pull request submissions, and I think we should include that in both documentation and automated AI-augmented tooling at the same time.

I'm more than happy to discuss any of the above, and I'm fully aware that this is currently a very controversial topic. But I think it's one where we need solutions sooner rather than later, because from exchanges with other open source maintainers, I know that some more trendy packages are currently really struggling with keeping up with AI-produced workload.

[swissspidy]

The good news is that we're not currently swamped with a ton of PRs. It's actually rather quiet in my opinion.

I'm hoping that with more guidance through "configuration" files like AGENTS.md and more automation we can increase velocity in the project while reducing maintenance burden. Some other things we can do:

• Reduce friction for new contributors when it comes to setting up the environments (e.g. when having to deal with MySQL/SQLite for Behat tests)
• Improve documentation and automation around things like coding standards and testing
• Have Copilot or similar do an initial PR review before pinging the team
• Improve new contributor experience, for example by posting a nice welcome message like WordPress core does

[wojsmol]

• Have Copilot or similar do an initial PR review before pinging the team

We can try coderabbit for PR reviews - there is free plan for open source.

[swissspidy]

I think we just use Copilot for now for reviews too because it seems to be free and is well-integrated.

[swissspidy]

Merging this one for now..let's see how it goes! We can always adjust or revert.