Beyond agentic coding
This piece critiques the prevailing 'agentic coding' paradigm, arguing it often hinders developer flow and productivity. Instead, it proposes integrating AI using 'calm technology' principles, advocating for subtle, enhancing tools over chat-based agents. The discussion on Hacker News largely resonates with this perspective, exploring AI's role in code review, team collaboration, and the nuanced impact on developer workflow.
The Lowdown
The author argues that current 'agentic coding' approaches, particularly chat-based LLMs, paradoxically decrease developer productivity and break flow states. This conclusion is drawn from personal experience, observed candidate performance in interviews, and research studies indicating increased idle time with such tools.
The article advocates for leveraging AI in software development through the lens of 'calm technology' design principles, which prioritize keeping the user in a continuous flow state. Key tenets include minimizing attention demands, making tools 'pass-through' to the actual work, and fostering a sense of calm. Examples of non-LLM calm tools like IDE inlay hints and file tree previews are presented to illustrate these concepts.
Critiques of chat-based agents highlight their high attention demands, indirect and slow interactions, and their tendency to undermine calm by requiring constant user engagement. Conversely, GitHub Copilot's inline and 'next edit' suggestions are examined as prior art, demonstrating how, with proper configuration, they can align with calm technology by being more 'pass-through' and less intrusive.
Finally, the author sketches out future AI-assisted calm technology concepts, such as:
- Facet-based project navigation: Semantically browsing code by intent rather than file paths.
- Automated commit refactoring: AI splitting large changes into smaller, reviewable commits.
- File lens: Features like 'Focus on…' to filter code by interest and 'Edit as…' to allow editing in a different language or format.
The piece concludes by urging developers to think beyond chatbots, asserting that chat is
The Gossip
Aiding Autonomous Auditing
The discussion extensively explores AI's potential to enhance human code review, rather than replace it entirely. Commenters suggest AI could automatically segment large diffs into more manageable, focused commits or generate comprehensive 'review plans' that logically order files. This approach aligns with the article's 'calm technology' principles by reducing the cognitive load on human reviewers, allowing them to focus on substantive changes while AI handles structural organization.
Flow State Feud: Chat's Challenge
A central debate revolves around the author's claim that chat-based agentic coding disrupts developer flow. Many agree, citing personal experiences of mental model desynchronization and the slow, indirect nature of chat interfaces. Others contend that chat agents are highly effective if used strategically for planning and iterative refinement, and that overly intrusive inline suggestions can be equally, if not more, disruptive than chat when not configured appropriately.
Interviewing with Intelligent Intermediaries
The author's observation that job candidates using agentic tools often performed worse in interviews sparked considerable debate. Some commenters agreed, suggesting that reliance on AI might lead to a shallow understanding of solutions or a lack of critical iteration. Conversely, others argued that if AI tools are integral to modern development workflows, interviews should adapt to test a candidate's proficiency in leveraging them effectively, drawing parallels to allowing Google or Stack Overflow in previous eras.
Collaborative Coding Conundrums
The conversation extends to AI's implications for team dynamics and collaboration. The 'junior dev' analogy for AI agents is a recurring theme, highlighting that while agents can accelerate implementation, humans still bear the burden of detailed review and maintaining architectural coherence. Concerns are raised about preserving a shared mental model within teams and building trust with AI. Some suggest AI could assist by generating design documents to improve clarity, while others note that Amdahl's Law implies AI's speedups are limited to only a fraction of the overall human-centric workflow.
Critiques and Calm Concepts
General reactions to the article and the concept of 'calm technology' varied. Many praised the article as insightful, appreciating its focus on subtle, workflow-enhancing tools. Others criticized it as overly generic, an unnecessary promotion of AI, or found the 'calm technology' concept somewhat naive for not addressing broader economic and social underpinnings of current tool design. The author also engaged directly, clarifying details about their website's naming and explicitly stating that AI was not used to generate the blog's content or layout.