Quoting Armin Ronacher

Armin Ronacher's latest essay, spotted by Simon Willison, zeroes in on something most discussions of AI coding agents skip entirely: the invisible glue that holds a codebase together isn't the code itself, it's the shared understanding among the humans who wrote it. Nobody writes that understanding down in full. It's scattered across doc comments, Slack threads, code review nitpicks, and the war stories people tell about why a service is shaped the way it is.

Before agents showed up, Ronacher argues, that shared understanding got maintained almost by accident, through friction. Wanting to touch someone else's storage layer meant reading their code, pestering them with questions, maybe looping in another team whose service quietly depended on some assumption nobody had written down. It was slow. Some of that slowness really was waste. But a chunk of it was the actual mechanism by which two engineers' mental models converged, or by which they discovered they'd drifted apart without realizing it.

That's the part Ronacher thinks people underestimate when they cheer on agents for moving fast and touching everything. An agent can rewrite your storage layer in minutes without ever having the conversation that used to force two humans to check whether they still agreed on how the system worked. Speed goes up. Synchronization goes down. Nobody notices the gap until two parts of the system quietly stop agreeing about what's true.

Ronacher isn't arguing against agents, he's flagging a cost that doesn't show up in velocity metrics. Codebases have always run on tacit knowledge as much as explicit rules, and removing the friction that used to transmit that knowledge doesn't make the knowledge unnecessary. It just means someone, or something, has to find a new way to keep everyone's understanding in sync, or watch the gaps accumulate until a change breaks an invariant nobody bothered to write down.