Four Levels of Agentic Programming (karimjedda.com)

In a recent discussion on the challenges of AI-assisted programming, a call for clearer classification in the use of AI tools by engineers was made to address the stigma surrounding so-called "vibe coding." The term, which has become a derogative label for AI-generated outputs, has led to a lack of differentiation in code contributions, making it difficult for reviewers to assess the true quality and originality of the work. This blanket labeling can discourage engineers who are diligently applying AI while still ensuring accountability for their code, ultimately stifling innovation and contributions. The proposed classification system categorizes code based on human accountability and AI assistance into four levels: A0 (fully human-written), A1 (human-written with AI snippets), A2 (agent-created but human-reviewed), and A3 (agent-created and unreviewed). By promoting self-labeling among contributors, teams can better ascertain the level of scrutiny required for each code submission and mitigate the risk of accumulating low-quality A3 code in environments that demand higher standards. This framework emphasizes the importance of understanding and explaining code, shifting the focus from the tools used to the quality and integrity of the code itself, thereby fostering a more constructive dialogue within the AI/ML community.