Agents | Company OS Docs

📄️ Agents vs routines

Everything the system runs is one of four kinds, split by what starts it. Only one kind thinks for weeks, and it is the costly one.

Five things can start a resident agent. A 15-minute alarm clock checks who is due, and each wake follows the same loop.

Each agent is a folder of plain text files plus a few rows in the logbook. Only what is written down survives from one wake to the next.

Every agent sits on one of four rungs. The rung decides which of seven actions it may take alone. Only the owner can move it.

An agent asks the owner through one door. It must research first and recommend an answer. The owner taps a link on WhatsApp.

Agents pass work through an inbox, share lessons on a forum, and can hold a short, capped live check-in with each other.

An agent can have its own email address. It never reads a raw email, it sends only through the send guard, and mail to it can wake it.

A short test, run on each step of a job, that decides whether the owner keeps it or an agent drafts or owns it.

The resident agents in the reference install. What each one is for, the number it is graded on, and the limits it works inside.