The core problem is simple: AI agents mimic the social surface of a colleague without any of the underlying properties that make colleagues trustworthy. They don't learn from your corrections, adapt to your context over time, or bear any consequence for bad outputs.
This matters because users are already responding to agents as if those properties exist. The social instincts that evolved for human interaction, turn-taking, tone-reading, reciprocity, get triggered by systems that are fundamentally stateless. Frustration follows when the mental model breaks against reality.
The full piece from Sidebar is worth reading for how it maps the specific gap between perceived and actual agent behavior. The argument is not that agents are useless. It is that the interface conventions we are building actively mislead users about what they are dealing with.
[READ ORIGINAL →]