Testing Astra's autonomy: it starts work on its own and oversteps
Given the same brief, the fast chat answered and stopped while GPT-6 Astra built a longer plan, reached for tools and went past the question, leading the author to adopt a one-chat-one-job rule.
The new GPT chat does not wait. That is the upgrade and the problem I ran the same brief twice. Once in the fast chat. Once in GPT-6 Astra Fast chat answered and stopped. Astra kept going. It opened a longer plan, reached for tools, and started building past the question I asked. The answer got better. The owner got weaker That is the point of the new version. It can carry a task to the end. It can also finish a task you did not start So I use it with the same rule I use for agents -> One chat. One job Write the stop in the first message: do not open extra tabs, do not publish, do not rewrite the brief If it starts building a product instead of answering, cut the run and take the keyboard Astra is useful when the work is long and you already know the owner. It is noisy when you only wanted a paragraph and it decides to ship a system The model got more agency. Your job is still to name who is allowed to continue