A decision you didn't write down isn't a decision

The most expensive failure in AI coding isn't a wrong decision. It's a decision that was made, then forgotten, and quietly unmade by the next generation. Derek Wang, author of Essay Three in AI Harness Engineering, outlines a raw law: a decision that lives only in conversation does not persist. It evaporates the moment the context window closes. If making it stick matters at all, it has to leave the chat and land in a file every session reads.

That file is the architecture layer of the pyramid—the ADR set. In Essay One, Wang provided the skeleton, and in Essay Two, the top. This essay covers the layer that stops drift no one notices until it costs a week. Wang has lived this failure too many times: after 40 exchanges with an agent, you lock in a hexagonal pattern where the domain layer never touches the database directly. However, it is locked in only as chat history, securing decisions about as well as a candle against a hurricane.

When the next session opens fresh, the model regenerates. Without knowing about the hexagonal choice, it wires the domain layer straight to a repository. Nobody spots the deviation until it spreads across twenty files, burying the original reasoning in unread logs. A conversation is not a contract; it is a one-time instruction stream with no history, archive, or power to bind the next session.

Stock photo for illustration only, not from the actual event

When decisions have nowhere to live, three distinct failures follow: Memory drift (humans remember, but the AI doesn't because it was never written down), Architecture drift (small innocent deviations compound over twenty regenerations), and Repeated-argument drift (re-debating the same question with every new agent).