Giving someone a bigger assignment changes what you need to explain. “Find me a few examples” can survive a vague brief. “Research the options and bring me a recommendation” needs a purpose, boundaries, and a way to judge the answer.
The same problem is arriving in AI work. This week’s model announcements promised systems that can carry more of an assignment across tools and files. An unusually concrete research project showed what it takes to trust the result.
The useful connection is between the size of the job and the strength of its finish line. As you watch fewer intermediate steps, you need clearer evidence at the end.
01 / THE ASSIGNMENT
Longer work changes the brief.
OpenAI introduced GPT-6 Astra around work that extends across computer tasks, documents, spreadsheets, and presentations. Those are the provider’s descriptions of the model’s intended capabilities, not a guarantee that a particular business process now works unattended. [1]
Consider a hypothetical supplier review. Finding three suppliers is a task. Comparing their suitability, documenting uncertainty, and preparing a recommendation is an assignment. The second requires decisions about which information matters and when evidence is too weak to proceed.
A system can produce a polished recommendation while quietly making those decisions for you. Before delegating the larger job, name the recipient, the decision, and the conditions that would make a recommendation premature. That gives a reviewer something more useful than “looks good” to work with.
02 / THE EVIDENCE
A finished artifact needs a separate test.
Anthropic described using Claude to formalize an existing proof of Fermat’s Last Theorem in Lean, a system that checks mathematical proofs. This was a formalization of established mathematics, not the discovery of a new proof. The research account also describes coordination difficulties and the value of shared project state. [2]
The practical lesson is the separation between producing an artifact and checking it. Most office work has no equivalent of a mathematical proof checker. You can still borrow the question: what could establish that this result meets the assignment without simply asking its author to reassure you?
For the supplier review, that might mean opening the original certification documents, reproducing a cost calculation, and checking the recommendation against requirements written before the research began. Each test is smaller than judging the entire document at once.
This is also where the analogy ends. A mechanically checked proof is a different kind of evidence from a persuasive commercial recommendation. Human judgment remains part of the latter, especially where preferences conflict or information is incomplete.
03 / THE STOPPING POINT
The finish line is not the only boundary.
Anthropic’s alignment and security update described measures intended to detect and interrupt concerning activity during tool use. It is a company account of its safeguards, with limits; it does not establish that any arbitrary workflow can safely run without supervision. [3]
That brings the assignment back to its middle. A good final check does little for an action that has already changed a record or contacted a supplier. Decide which steps may happen during the research and which require a person.
For our hypothetical brief, reading approved sources and drafting comparisons may be reasonable. Sending a commitment to a supplier is a separate decision. Keeping those permissions distinct lets you experiment with a more complete assignment without treating every possible action as part of the same permission.
04 / PUT IT INTO PRACTICE
Write the finish line first.
Choose one recurring research or preparation task. Before the next run, write the checks a colleague could use to accept the result. Then give the system the task and those checks together.
Keep the trial small enough to review in full. Record which checks passed, what had to be corrected, and whether the larger assignment reduced your total effort.
One assignment. An observable finish.
Build: A task brief and acceptance checklist for one delegated assignment.
Look for: A reviewer can check the important claims and identify missing evidence; compare the complete effort with the current approach.
Give the system enough room to do useful work, and give the reviewer enough evidence to decide whether it did.