Coding agents quietly erode a team's shared understanding

Armin Ronacher argues that coding agents are quietly changing how software teams fall apart. His starting point is friction. Before agents, making a change in an unfamiliar part of a system meant talking to the people who owned it: a code review, a design conversation, a back and forth that was slow but did real work. It kept everyone's mental model of the system roughly in sync, and it forced each developer to learn the parts around their own.

Agents remove that step. A developer can now ask an agent to add OAuth, wire up a caching layer, or reshape a database schema in code they have never read and do not understand. Each change compiles. Each passes its tests. Taken together, across many developers working this way, the architecture drifts and no single person holds a coherent picture of it. Ronacher borrows Bruegel's painting of the Tower of Babel: in the story, construction stops once people can no longer understand each other. With agents, the tower keeps rising anyway. The code keeps working while the shared understanding underneath it decays, which is the more dangerous failure because nothing visibly breaks to warn you.

Why it matters

If you lead an engineering team, the productivity from agents can hide a growing gap in who actually understands your system. That points to keeping humans in the loop for architectural decisions and design review, not only for the changes that happen to fail their tests.