Ask HN: Are you getting addicted to the dev workflow of coding with agents?

Why Read This from an HCI Perspective

This piece compares the experience of using LLMs (large language models) to ‘uncertain rewards’ and a ‘Skinner box,’ explaining why users keep repeating prompts. For HCI/UX practitioners and researchers, it’s meaningful because it shows that generative AI interaction is not just an efficiency tool—it’s a behavior design problem intertwined with expectations, rewards, and learning. It also explores how agentic tools may change users’ immersion and reliance.

CIT's Commentary

This comment is interesting because it captures the core value of generative AI not in ‘generating correct answers,’ but in ‘designing the reward structure.’ A system where the output changes every time you submit a prompt can help exploration, but it can also create reinforcement-learning-like interactions that pull users into repeated runs. That’s why UX evaluation shouldn’t focus only on accuracy or speed; it should also consider expectation management, tolerance for failure, and the ability to stop. The second comment’s ‘future metaverse’ analogy may sound exaggerated, but it can be read as a signal that an interface wrapping around the work process could become a new computing paradigm. In the end, the key isn’t fantasy—it’s how predictable and controllable the collaboration partner becomes in everyday work.

Questions to Consider While Reading

• Q.In LLM interactions, do ‘uncertain rewards’ increase user engagement, or do they lead to fatigue and overuse?
• Q.When evaluating the UX of generative AI, what behavioral and qualitative indicators should be considered alongside accuracy?
• Q.As agentic tools move deeper into the workflow, how should users’ sense of control and responsibility be designed?