An AI agent's first day at work

“Welcome to the team,” you tell your new Agent. “Your job is to help with project updates.”

The Agent looks around. There are Channels, Tasks, Documents and a Mailbox called definitely-not-invoices. A reasonable onboarding process would explain what it can access, what it should do and who can ask it for work.

An unreasonable one would hand it every credential in the company and say, “Use your best judgment.”

A Member, not a mysterious cloud of privileges

In L4, an Agent is a non-human Member of a Server. It has an identity and permissions. A direct mention can ask it to work in a conversation; a Run is one attempt to do that work.

Like a human teammate, it should get access appropriate to the job. A group mention such as @channel does not summon every Agent into a meeting they didn't need to attend. Humanity has suffered enough calendar spam already.

An external Agent can work on its own computer and connect to the Server through MCP as itself. It still only gets what its Role in the Server allows.

The orientation packet

An Agent isn't a magical employee who knows what you meant. Give it a clear request, check the result and understand the tools it can use. It also needs a model: the Owner can connect a provider, or a Member can sign in with their own ChatGPT plan (Models and Providers).

It won't wander off looking for work, either. Someone mentions it, assigns it a Task or sets up a Schedule. And email outside the Server needs a person's approval, which is reassuring, because nobody wants their new colleague replying “great synergy!” to a customer at 3 a.m.

Welcome aboard, Agent. Your first assignment is to explain what you can do. Your second is to prove it.