How the agents work each day, and the fixes I had to make by hand

«Part 27»

I wanted a team that would work without me poking it all day. In the last part I talked about what was already in place. The practical problem now is different. Creating an agent is not the job. I need them to run on their own, every day. And when the house falls down, I need to know what to fix with my hands.

An agent is not magic. It is routine, memory, and a written rule. If one of those three goes missing, the team stops.

What an agent’s day actually is

SOUL is the identity card: who the agent is, what it may do, what it may not. Memory is what it keeps from one day to the next. The gateway is the door. When the notebook boots, that door lets the agent talk to the world: chat, calendar, tools. Scheduling is the alarm clock. In our case, 07:47.

I do not send reminders all day. The rule is to work on automatic. Nothing goes live without my “aprovo”.

Here is what I used. Hermes on the notebook, with agent profiles. SOULs with written rules. A gateway set to start at Windows logon, when I turn the machine on. A local 07:47 job for picking topics and starting the newsroom. A company notebook, with date, agent, and subject. Me, Rodrigo, on final approval.

In the morning the 07:47 alarm starts topic selection. Then the pipeline runs: reporting, draft, art, fact check, and the editor, who returns a humanized draft. One rule I will not drop: the same agent does not write a piece and then check its own piece.

I only show up to read and say “aprovo” or send it back.

Outside the newsroom, other roles follow the same logic. Clear role. Do not reread old chat. Fix the team without waiting to be asked. Automatic, here, means delegate with a lock. It does not mean walk away and hope.

What the reinstall wiped

After I reinstalled Windows on the notebook, the environment came back clean. Empty memory. The schedule was gone. The gateway did not start on its own. That is when I saw the gap between a daily automatic run and the hand fixes that put the system back on the rails.

The computer came back from zero. What looked already configured did not travel with it. Reinstalling the system is cleaning the house, including what the agents needed in order to exist.

Without the door open, the agent still exists as a profile, but it does not enter the day. I made the gateway start when I log into Windows, so I do not have to open it by hand. If it depends on me remembering, it is not automatic.

The newsroom alarm vanished with the reinstall. I created the 07:47 time again. A routine without an alarm is just a wish.

The agents woke up with no history. They do not remember me just because the program is installed. I accepted the empty state, wrote the essential bits again, and did not pretend the past was still there.

Whatever lived only in old conversation was gone. I wrote it into the SOULs again: work on automatic, do not reread old chat, nothing goes live without “aprovo”, language, the company notebook, a short English tip at the end of a message to me. A rule that is not in the SOUL does not survive a reinstall.

What I would do differently

Treat SOULs and the alarm as part of the install, not as “I’ll get to that later”. Keep a short recovery checklist: clean Windows, gateway at logon, 07:47, SOULs, essential memory. Do not leave an important rule only in that day’s chat.

This is for people who run agents on their own computer and want a routine, not a one-day toy. It is not for anyone who expects the program to remember everything after a format.

The team works on its own because the rule is written down and the alarm exists. The reinstall showed what was real and what was only habit.

What goes into the permanent recovery checklist is still open. I pick that in the next part.

Sources

• Hermes Agent docs: https://hermes-agent.nousresearch.com/docs