AGENTS.md as a dark signal (joshmock.com)

In a recent discussion, a senior engineer reflects on the transformative impact of AI, particularly large language models (LLMs), on software engineering practices. While acknowledging the ambivalence around AI's role in enhancing productivity, they express concern over its implications for job markets and code quality. A notable experiment involved using GitHub's Copilot agents to automate long-standing coding tasks. However, the engineer highlights common pitfalls, such as agents generating unit tests that fail due to overlooked issues, emphasizing the need for more robust contextual memory in AI tools. As a solution, the engineer introduces the concept of an AGENTS.md file to document learnings and insights from AI agents, proposing that it could serve as a resource for future AI-driven coding tasks. This practice could mitigate risks associated with haphazard AI code contributions, perceived as “vibe coding.” While the presence of such files might raise concerns about overall code quality for some seasoned developers, it also acknowledges the reality that many developers are unconsciously incorporating AI suggestions in their work. Thus, the AGENTS.md could play a crucial role in maintaining oversight while adapting to the growing influence of AI in software development.