December 30, 2025
Leash the bots, unleash the drama
Prof. Software Developers Don't Vibe, They Control: AI Agent Coding Use in 2025
Devs say AI needs a leash, not a cheer squad — and sample sizes get roasted
TLDR: The paper says pros use AI coding assistants for speed but keep tight control, treating them like tools, not bosses. Comments clap back over tiny samples, hype fatigue, and messy untested code, with a popular view that real skill is catching AI’s confident mistakes — and fencing them in.
A new study says experienced coders like AI “agents” (software helpers) for speed but refuse to hand them the steering wheel. The community didn’t just nod; they pounced. One camp cheered the framing that pros don’t “write code,” they steer systems, using AI like power tools while keeping human quality control. Another camp rolled their eyes at the study’s tiny sample — N=13 observations and N=99 surveys — with skeptics calling it soft science and demanding harder data. Cue the popcorn.
The thread’s mood swung between pragmatic and spicy. A top comment nailed the vibe: the real skill isn’t fancy prompts, it’s knowing when the bot is confidently wrong — and fencing it in. Meanwhile, testers cried foul that 53 people used agents to build apps vs just 1 for testing, joking that AI code arrives like a party nobody cleaned up. One poet likened good dev work to “maintaining fertile soil,” while another snapped, “I. Don’t. Care.” after years of hype. The drama is clear: AI is a speed boost, not a boss, and devs are here to control it — leash on, eyes open. For context on “agents,” think smart assistants that write and edit code, but still need a human editor.
Key Points
- •Study investigates how experienced developers use AI agents in software development.
- •Methods include field observations (N=13) and qualitative surveys (N=99).
- •Developers view agents as productivity boosters but maintain control over design and implementation.
- •Emphasis on fundamental software quality attributes drives strategies to control agent behavior.
- •Paper identifies suitable task types for agents and calls for improved agentic interfaces and usage guidelines.