assistant

In depth

Working turn by turn

An assistant works in turns: you write, the harness assembles what the model has to read, the model produces an answer, the exchange stops. Between two turns nothing keeps running, no process pursues a goal in the background. The impression of a conversation comes from the harness, which puts the history back into the context at each request, and not from the model, which keeps nothing. This is why a long conversation ends up losing its beginning: it is not forgetting, it is an overflow of what can be given back to read.

Three neighbours

Three neighbours deserve to be told apart. Chatbot names the interface, the chat window: an assistant can do perfectly well without one, for instance when it lives inside a word processor or a code editor. The agent receives a goal and chains the steps without coming back to you: the boundary is not the richness of the tools, it is the initiative. Between the two, many products offer adjustable autonomy, where the user authorises a run of steps before taking control back, and naming that intermediate zone spares a good many fruitless debates.

What gets measured

An assistant is judged on the answer, not on the task: what is measured is accuracy, the citing of sources, the useful refusal, the delay. The recurring trap is to credit the model with what the product owes to its harness, then to conclude that the model must be changed. The second is to confuse fluency with reliability: an assistant always writes something, including when it holds no relevant information. The third is organisational: because it is easy to reach, an assistant spreads through a company without anyone having decided what data may go into it.