GPT-6 Astra and the work you no longer see being done

By Nate from Nate's Substack · September 6, 2026.

The briefing's starting point is a simple idea: throughout the history of computing, the machine waited for the human to decide what to do and when. It performed that specific task and stopped. Nate argues that this relationship began to break down in 2025, when agents became able to take an objective and work toward it continuously. Over the past year, he says, his work has consisted of setting objectives for systems such as Sol, Opus 5 and GLM 5.3, and then judging what they returned.

What changes now, according to the author, is this week's launch of GPT-6 Astra by OpenAI, which he credits with transforming "everything that happens in the middle" of the work process. His argument is that computing has become "self-directed": it is enough to state an objective briefly and imprecisely for the machine to work for days without supervision and without going off the rails, delivering at the end a finished piece of work that incorporates hundreds of decisions the person never witnessed.

The consequence Nate underlines is one of responsibility: the user ends up answering for decisions he never made himself, and his role becomes spotting the ones that went wrong before anyone else notices.

The email works as a summary/preview of a longer written "companion" (with seven key points) that complements a video, and it lists four of those central ideas:

First, that "the middle" of a person's work disappears: what remains is a brief specification at the start of the task and a judgment or assessment at the end.

Second, that large companies also lose their own "middle layer": that organizational layer which existed to tie up loose ends is no longer necessary, which in turn allows small companies to become far more capable without needing that intermediate structure.

Third, a legal argument: the user is responsible for what his agent commits to on his behalf. Nate points out that a body of law already exists governing people who act on behalf of others, and argues that this legal framework already applies to machines acting by delegation as well.

Fourth, that the last two percent of reliability is worth trillions of dollars: a system's reliability determines which tasks people are actually willing to delegate to it, and the gap between 98% and 100% reliability is, according to the author, where the business value of these technologies is concentrated.

The email notes that the full briefing, available at the link, develops the seven key points (only four of which are detailed here) and explains what each one changes in the reader's day-to-day. It closes with a call to subscribe to the "Executive Circle" tier, which gives access to these additional Sunday briefings and to the author's own MCP server.