The most transferable demo of launch week was not a coding benchmark. Ethan Mollick pointed GPT-6 at tens of thousands of his own emails, writings and calendar entries and had it assemble a personal knowledge base of research, contacts, ideas, relationships and tasks.
The most transferable demo of the week was not a coding benchmark. Ethan Mollick pointed Astra at tens of thousands of his own emails, writings and calendar entries and had it assemble a personal knowledge base of research, contacts, ideas, relationships and tasks.
That is a company-shaped job wearing a personal disguise. Every organisation has a decade of decisions sitting in inboxes, documents and calendars that nobody can query, and until now assembling it was a project nobody would fund. Long-horizon computer use changes the arithmetic, because the model can grind through the corpus for hours without supervision and the cost is tokens rather than a team.
If you do this, do it on a copy, scope credentials to read-only, and pick a corpus with a clear boundary. The failure mode is not a bad summary. It is a model with live credentials to everything because it was easier than scoping.
Two constraints make the difference between an asset and an incident. Scope credentials to read-only and work on a copy, because a long-horizon run against live systems is the worst possible place to discover a permissions mistake. And pick a corpus with an edge: one mailbox, one function, one date range. A boundary you can describe in a sentence is a boundary you can audit.
The output is worth more than the summary it produces. What you are actually building is the queryable layer that your AI does not currently have, assembled from the decisions your company already made and then forgot.
Why the personal version is the company version
Mollick's run is a personal knowledge base, and that is the disguise. The same job at company scale is the thing every AI project quietly assumes exists and then discovers does not: an authoritative, queryable account of what the organisation knows and why it decided things.
Systems of record capture what happened and almost never why. A CRM records that a deal slipped; it does not record the judgment behind the forecast. Decision rationale is among the most perishable assets a company owns, because it lives in Slack threads, escalation calls and the heads of people who leave.
What long-horizon computer use changes is that reconstructing it stopped being a funded project and became an overnight job. Not because the model got smarter at reading, but because it can now read for eight hours without a human sitting next to it.
Bound it before you run it
One corpus, one edge. A mailbox, a function, a date range. If you cannot describe the boundary in a sentence, you cannot audit what the agent touched.
Read-only, on a copy. A long unattended pass against live systems is the worst possible place to discover a permissions mistake, and the cost of a snapshot is trivial next to the cost of a write you did not intend.
Test the revoke path first. Gergely Orosz named the pattern this summer: the connect action grants more than the user believes and the disconnect action revokes less. Claire Vo cut an assistant's Google access at 11am and still received an emailed summary at 2pm. Before a real corpus, prove you can turn it off.
Decide what the output is. A summary nobody queries is a document. The asset is a structured, addressable layer that later agents read from, which is the difference between a report and infrastructure.
The 922,000-token input ceiling is what makes any of this practical, and it is also the constraint: a corpus larger than that has to be chunked and compacted, which is the same discipline that bought the 37 points in Part 2.
What decade of decisions can nobody in your company query?
Book a free Diagnostic: 30 to 45 minutes, no deck, no pitch. We pick one bounded corpus and scope what a read-only overnight pass would actually produce.
Book the Diagnostic →Part 5 of Working with GPT-6 Astra: The Operator Guide.
