Claude Code, OpenAI Codex and other agentic coding tools have a problem with .. “agency”.
I mean user agency of course.
Quick primer – these are “coding agents”. Claude Code runs on the command line and OpenAI Codex works by checking out your GitHub repo in a cloud environment.
You give these tools a prompt and they reason, plan and execute changes. They may ask some questions but typically there is minimal human intervention.
My experience with using these tools has been similar to working with the most annoying kind of junior developer. Extremely confident, asking very few questions, and generating code that is sometimes useful and sometimes terrible. And since model capabilities have a “jagged frontier” – it’s not always clear if the code will be 💐 or 💩.
This is a terrible user experience!
The agentic implementations in Cursor and GitHub Copilot are much better because the “diff -> approve” workflow allows the developer to accept or reject specific changes.
So – software agents are great, but if they keep removing agency from the humans they will see significant barriers to adoption.