We counted eleven agents in one company of eighty people. Four had been built by someone who had since changed teams. Two sent email. One of those still had access to a shared inbox nobody monitored. The inventory existed because IT had finally asked. The jobs did not, because nobody had required a job before the thing was allowed to run.
A directory of confusing tools is still a list of confusing tools.
a job is a boundary
Write it the way you would write a role, not a prompt. This agent drafts the weekly status from these three sources. It does not send the status. It does not touch billing. When a source is missing, it stops and says so. Maya owns it. If Maya leaves, the agent pauses until someone else signs.
That paragraph is the product spec. The prompt is an implementation detail.
If you cannot say what it must not do, you do not have an agent. You have a permission.
then keep fewer of them
Most of the value is in a short list of agents people actually rely on, each with a job and a failure plan. The long tail is leftover access. Shut it off. The teams that get something out of this are not the ones with the biggest catalog. They are the ones who can explain any item on the catalog without opening the repo.
Add an agent to the list when it has a job, an owner, and a way to stop. Not before.