meta: add a root-level AGENTS.md

[jasnell]

We're seeing a number of repeated mistakes in AI-generated contributions, including agents using Signed-off-by, creating what appear to be unsupervised PRs, appearing to respond to comments without human oversight, etc. Having an AGENTS.md file with specific instructions to the agent on what to avoid should help.

[legendecas]

There is an ealier PR for AGENTS.md: #65463

[MikeMcC399]

Good to see the experience of receiving and responding to AI-assisted / -generated PRs has resulted in a proposal for explicit coverage in an AGENTS.md directive!

#62612 disabled linting for AGENTS.md. Consider re-enabling, then re-formatting and responding to linting findings. I noticed because #65463 from @BethGriggs had done this.

[joyeecheung]

The document uses a tone as if it's speaking to a human, even though it is not supposed to be read by human? e.g. the distinction of collaborators or non-collaborators is pointless to an agent unless you give it permisssions where the distinction makes a difference, but then the document goes on saying the agent shouldn't use any of these permissions, which means the agent is not supposed to make use of that distinction in the first place.

[jasnell]

The document uses a tone as if it's speaking to a human, even though it is not supposed to be read by human? e.g. the distinction of collaborators or non-collaborators is pointless to an agent unless you give it permisssions where the distinction makes a difference, but then the document goes on saying the agent shouldn't use any of these permissions, which means the agent is not supposed to make use of that distinction in the first place.

Agents are perfectly capable of correctly interpretting it as written.