FedEx has a rule for AI agents: even the ones its own employees build must be registered on the company's platform, and every action an agent takes must be traceable to a human. Its chief executive also wants half of the company's workflows run by agents as soon as possible. Those two goals are connected, and the connection says something about what happens to anyone who builds an agent at work.
A registry is what lets a company say yes
A company cannot hand half its work to things it cannot list. The questions a board asks are simple: what is the agent doing, and if a customer or a regulator asks, can we explain it? An agent nobody registered has no answer to either. So the register is not a brake on the plan to use more agents. It is the condition for it, the same way Capital One built its governance before it deployed agents at all.
It also changes who gets trusted with building. When a company decides who may create agents, it is deciding who it will let answer for them, which is why the people who know the work best, as in Uber's 30,000, are the natural builders.
Spreadsheets show how this goes
The lead example was the Sarbanes-Oxley law and the spreadsheets it pulled into audits. The part worth taking further is what spreadsheets looked like before that. Small systems built by end users are usually used by one or two people, rarely documented, and often hold copies of company data. Two of them can disagree about the same figure, and nobody knows which is right. Registering them fixed that without taking them away from the people who built them.
Agents need the register more than spreadsheets did. A spreadsheet with a wrong formula sits there until someone reads it. An agent with a wrong instruction sends the email, changes the booking or moves the order before anyone looks. Software makers are building the same rule into their products: Microsoft's system for company agent identities requires every agent to have a human sponsor, who decides when the agent is no longer needed.
What it looks like for the builder
Take a warehouse shift manager who builds an agent that reads carrier delay emails, checks each against the open orders and drafts the message to the customer. If a shift handles 40 of those emails and each takes four minutes by hand, the agent gives back close to three hours a shift. Under a FedEx-style rule, the manager registers it, and the log shows every email it drafted and every order it looked at. The manager also decides which messages it may send alone and which wait for approval, for example anything that promises a refund.
That is more responsibility than a spreadsheet ever put on its author. It is also a promotion of sorts. A company that wants half its workflows run by agents will need a very large number of people who know a process well enough to put their name on a machine doing it. Building the agent is the easy part. Being able to explain, in front of a customer or a regulator, why it did what it did is the skill that will be scarce.