AGENTS.md: Align AI attribution policy with Kubernetes community#204
AGENTS.md: Align AI attribution policy with Kubernetes community#204cgwalters wants to merge 1 commit into
Conversation
There was a problem hiding this comment.
Pull request overview
Updates AGENTS.md to align this repo’s AI-attribution guidance with the Kubernetes community approach: disclose AI assistance in the PR description (not via commit trailers), and adds additional guidance aimed at improving reviewability and human ownership of prose.
Changes:
- Replaces the prior “Assisted-by” commit-trailer recommendation with a policy to disclose AI assistance in the PR description and explicitly prohibits AI credit trailers.
- Adds guidance encouraging smaller PRs by splitting preparatory work into separately reviewable PRs.
- Adds guidance encouraging humans to author/verify commit-message prose and review other natural-language text for accuracy.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Adopt the policy established in kubernetes/community#8918: disclose AI assistance in the PR description rather than via commit trailers. Listing AI tooling as a co-author or using 'Assisted-by' / 'Co-developed-by' trailers is now explicitly prohibited, matching the Kubernetes steering committee's position on AI disclosure. Further rationale: the Assisted-by schema (tool + model name) doesn't capture common multi-agent patterns (planner + executor), and model names alone don't tell reviewers much — the prompt harness and skills matter just as much. Also adds two new guidance sections: - Pull request size: encourage splitting prep commits out separately - Commit messages and text: remind agents that natural language can't be machine-checked, so humans should own the commit message prose Signed-off-by: Colin Walters <walters@verbum.org>
cgwalters
force-pushed
the
sync-agentsmd-k8s
branch
from
02b4de1 to
55be9b7
Compare
jmarrero
left a comment
jmarrero
left a comment
There was a problem hiding this comment.
This is OK with me. My first time reading this it felt kind of hostile to agentic development. But thinking about this, even when I am letting the agents do all the code I spend time making sure it's tested and is doing what I think it's doing. It is not good to have the reviewer be the tester or first person that looks at the code we submit specially if generated by llms.
|
Hmm what felt hostile? Can you elaborate? That was definitely not the intent. |
|
Maybe hostile is not the correct word, maybe discouraging of fully automated agentic development? To be clear I don't think the text is hostile towards the human reader. And it's not our policy here but the Kubernetes one as well. If we were to adopt a fully automated flow with loops that check issues, create the fix, test, and push the PRs then these PRs probably won’t be taken seriously as the human only set-up the loop. A human would not have fully reviewed or tested the code. Or even accepted based on the human requirement for the commit message and taking it out of draft. I agree that we need to guard against a crazy amount of low-quality PRs to review and merge which is why I approved. I could just be buying too much the hype of "loop engineering" some AI devs are selling online. |
Adopt the policy established in kubernetes/community#8918: disclose AI assistance in the PR description rather than via commit trailers. Listing AI tooling as a co-author or using 'Assisted-by' / 'Co-developed-by' trailers is now explicitly prohibited, matching the Kubernetes steering committee's position on AI disclosure.
Further rationale: the Assisted-by schema (tool + model name) doesn't capture common multi-agent patterns (planner + executor), and model names alone don't tell reviewers much — the prompt harness and skills matter just as much.
Also adds two new guidance sections: