The Agent Needs a Manager

OpenAI wants agents to become normal work software, not a coding-side phenomenon.

Good. That is where this was always going.

But the hard part is not access. It is management.

Codex worked first because code already has rails: diffs, tests, owners, branches, rollbacks. The agent can produce unfinished work without the team pretending unfinished work is final.

Most office work does not have that shape.

A sales follow-up, board deck, vendor comparison, or weekly operating brief depends on context scattered across docs, meetings, Slack, CRM fields, and somebody’s memory. Give an agent that job without an owner and you do not get a coworker. You get an intern with file access.

That is the adoption problem hiding under the product launch.

Who gives the agent the goal? What can it read? What can it change? Who reviews the output? What counts as done?

If a team cannot answer those questions, a better model will not fix the workflow.

OpenAI can put agents everywhere.

The business still has to decide who manages them.