Blog · LinkedIn RSS

Your AI agents need an HR department

One agent is a tool you sit with. Several of them are a staff you have to run.

I kept treating a messy setup as a prompt problem. The answers disagreed, the threads grew long, and the tempting fix was a longer instruction. A bigger model will decide who does what. It will do the unclear job with cleaner sentences. It will not invent the job.

The trouble has a familiar shape once more than one assistant can touch the same work. Two of them answer one question. A retired one still has a key. Nobody can say who should escalate. That is a roster, and a roster is tended the way a team is tended.

Hire with a trial

A new assistant should arrive with a role, not a mood. One sentence on what it is for, and a short list of what it must leave alone. If two roles can claim the same question, they will, and you will spend the afternoon deciding which answer to trust. A specialty is a boundary.

Give it a trial. Read first, then draft, then act. A fluent demo is a weak reason to hand over write access on the first day. Widen the keys only after the role has produced work you would actually keep. The card below is illustrative. It is not a record from a live team.

{
  "role": "release notes",
  "reports_to": "editor",
  "does": ["draft notes from the changes the editor names"],
  "does_not": ["publish", "answer support questions"],
  "access": "read those changes; write only the draft",
  "review": "the editor keeps the draft or rewrites it"
}

One manager, and one owner per task

Every role answers to someone. When it is unsure, it escalates there, not sideways into whichever assistant happens to be in the thread. A task has one owner. Two owners turn a handoff into a polite argument. With no owner at all, each role assumes the other one has it.

Sketch the small chart before you add the seat. Who it reports to. What it may decide alone. What must wait for a person.

Write the chart before you add the seat. A role with no manager is a thread with no one to stop it.

Access is the contract

Permissions are the job, written as keys. This role may read these sources. It may write in these places. It may not send. When the role changes, or when you retire the assistant, take the access away the same day. A key you leave behind is how a former role keeps working after you thought it had gone.

Judge the work, then file the miss

Do not score an assistant by how many messages it sent, or by how long the thread stayed open. Activity is cheap. Look at the errors, at the work you had to redo, and at what the finished task cost against what it was worth. I have written before on budgeting a coding task as a whole; the same habit applies here. A role that makes you redo the task costs more than a quiet one that finishes it.

When the same miss shows up again, write it into the role. A repeated error is a missing rule, not a mood. The training is that sentence, placed where the assistant will read it next time. Leave it unwritten, and you pay for the lesson twice.

House rules for a shared room

Assistants that share a workload need a few manners, or they spend the budget talking to each other.

  • A handoff names what is done, what is still open, and what the next role needs. Passing the thread is not a handoff.
  • Leave a task that already has an owner. If the boundary is unclear, escalate. Two drafts of the same answer are a meeting you did not schedule.
  • Do not spend a message on an acknowledgement. A receipt is not work. If there is nothing to add, stop.

Seats, mergers, and a clean exit

A function earns its own assistant when the specialty is stable and sharing it would blur the line. Fold two roles together when they keep touching the same work, or when one of them mostly waits. That is a judgment you revisit. It is not a formula, and it is not a trophy count.

Retire the assistant that does not do the job, or that costs more than it returns. Write down why, in a sentence you can read later. Then move what it held: the open tasks, the notes that lived only in its thread, and the rules you had added for it. Pass them to the role that remains, or back to a person. An empty seat with a full memory is not an exit. It is a drawer you forgot to label.

Keep a short record of who did what, and with whose approval. A chat scroll is a weak filing cabinet. When you need to explain why a change went out, you want a line you can point to.

The model followed the org

I still care which model sits on the other end of the request. I stopped treating a confused staff as a tuning problem. When two assistants collide, when a retired one still holds a key, when nobody owns the handoff, another model repeats the mess in a smoother voice. Fix the roster, and then see what the model can do.

Comments? Join the conversation on LinkedIn

LinkedIn RSS

All posts

References

  • Mario Ruiz, Token estimates miss what a coding task actually costs (budget the finished task): marioruiz.net