What the newsroom taught me about agents

I learned agents first in a newsroom, not in a lab. The work was obvious: research a company, pull a filing, draft a brief, assemble a package a desk could use. Software could do that at real volume. The thing it could not do — the thing it must not do — is publish.

An editor still said yes. That was not a cultural leftover. A market-moving sentence on Moneycontrol is a write with a name on it. The organisation already had a gate. We did not have to invent the idea of approval. We had to keep it once the draft got fast.

What the current rooms get wrong

They have the opposite problem. A working prototype, and no one who will sign for a write. They treat that as a change-management issue, or a model-quality issue. It is a missing specification. Until someone will answer, in writing, what may be drafted and what may be executed, the prototype is waiting for permission it will never get.

More tooling does not move that. A better model does not move that.

What I kept

Speed up the draft. Do not speed up the publish. The record of who said yes has to be as easy to find as the output. Undo has to be possible the same day, because a newsroom already knows what a correction costs.

If your agent can send and you cannot name the editor, you do not have a production agent. You have a draft that left the building. I would rather be accused of going slowly at the write than of inventing a new way to be wrong in public.