An AI agent working for three days needs more than a good initial instruction. It needs a human who can see where the work stands, resolve a blocked decision, and decide whether its result is ready to use. OpenAI’s October 2 guide to the GPT-6 family says agents can now take on tasks spanning hours or days and discusses how to manage that long-running work. For a business leader, the first question is not which model can stay busy longest. It is who owns Tuesday when the assignment began Monday and the result is due Wednesday.
TL;DR
Treat a multi-day agent assignment as a piece of shared work, not a request you leave in a chat window. Name a human owner, define what the agent may do without asking, require visible progress at agreed points, and make approval of the final result a separate decision. If nobody has time to review the work in progress, the task is not ready to run unattended.
The middle of the job is where ownership disappears
Consider a team preparing a response to a vendor renewal. The agent can collect contract terms, compare them with invoices, draft a list of discrepancies, and propose questions for the vendor. It may take several passes because the documents live in different places and a department head has to clarify which services are still used.
On Monday, the finance lead assigns the work. By Tuesday afternoon, the agent has found a price increase but cannot tell whether an old service was deliberately retained. It can keep going and make an assumption. It can stop and ask. Or it can quietly produce a plausible comparison that treats the service as waste. Only one of those paths may be appropriate, and the answer depends on information a person holds.
That is not a model-selection question. The finance lead needs to know what the agent has checked, what remains unknown, and which assumption would change the recommendation. A task can be continuously active while the underlying business decision is stalled.
The distinction matters more as people start running several agents at once. In OpenAI’s June 2026 study of Codex use, users at the 99th percentile regularly generated more than 60 hours of agent turns in a day across parallel agents. That is a measure of agent activity among unusually heavy coding-tool users, not 60 hours of human productivity or evidence that every business should operate this way. It does show why a single person may soon supervise more work in progress than they can remember from the chat history alone.
Give the assignment a working agreement
Before starting a long task, I would ask its owner to write down four things in ordinary language. What finished work looks like. Which sources the agent is allowed to use. Which decisions it can make without approval. And when a person needs to see a progress report. These fit in a short brief; they do not require a new management platform.
For the renewal, the agreement might say: compare the current contract with twelve months of invoices; list price changes and unused line items; do not contact the vendor or alter any records; stop and ask if service ownership is unclear. By Tuesday at noon, return a table of confirmed findings, unresolved questions, and documents still needed. By Wednesday, produce a draft recommendation with links to the source documents.
The noon checkpoint is not a ceremonial update. It is the moment the finance lead can tell the agent, “That service belongs to a team we acquired; ask their director before you call it unused.” Without that correction, the Wednesday recommendation may be polished and wrong.
Give the agent permission to flag a bad assignment, too. If the invoices are incomplete, it should say so rather than make the comparison look complete. A collaborator who cannot report uncertainty is a poor collaborator, however quickly it writes.
Progress is not the same as permission
A status report should let the owner answer three questions without opening every source file: What has been checked? What is blocked? What decision is next? A claim such as “80% complete” is less useful than “all invoices reviewed; two disputed line items need an owner.” Completion percentages can hide the one unresolved issue that changes the outcome.
Keep execution permissions separate from review. The agent may be able to read approved files and draft a recommendation. That does not mean it should email the vendor, change a payment, or circulate the recommendation as a final position. A named person owns those boundaries. If the work crosses a legal, financial, or customer commitment, that person should be able to see the evidence before approving the next action.
For a first trial, use one real assignment with a deadline that permits intervention. Record how often the agent has to ask for clarification, how much time a human spends reviewing, and which errors are caught before the result is used. Compare that with the team’s usual way of doing the same work. If the agent saves drafting time but creates a new pile of unreviewed decisions, it has moved the bottleneck rather than removed it.
The useful change is a shared rhythm: the agent does the legwork, the human resolves ambiguity, and the record of each decision travels with the task. That is a better picture of an agentic organization than a fleet of agents running without anyone knowing what they are doing. When an assignment lasts three days, Tuesday is where collaboration either becomes real or falls apart.
Research and structure: Mai. Editorial direction and voice: John Lipe.