Hand-overs are where things bend
More and more work is done by AI agents passing tasks to each other: one reads, one plans, one writes, one checks. Every hand-over is a retelling. Each agent adds a little of itself, drops a little it thought was unimportant, and the next one cannot tell which parts are original. No single step looks wrong. The end is still wrong.
What careful mode changes
In careful mode each listener says the sentence back before passing it on, and the sender compares. If too many words changed, the listener is told and tries again. That one cheap habit stops most of the drift, at the price of twice the messages. Notice what still slips through: a check that counts words lets a changed number walk straight past. Checks only catch what they look for.
What is simplified here
The listeners are one small model, SmolLM2-360M-Instruct, about 360 million parameters and a 276 MB download, given ten different instructions. It is tiny next to the models behind real agents, and it is told to bend the sentence on purpose; real agents drift by accident and far more slowly. The drift meter counts surviving words, not meaning. When the model writes something unusable, a hand-written rule fills in, and the list of the whole line marks every sentence that came from one.
Everything runs in this tab. Your sentence is never sent anywhere; the only thing fetched is the model itself, once, from Hugging Face, and then your browser keeps it. If you share a chain, its sentences travel inside the link you send and nowhere else.