Getting started
Four steps. Ten minutes.
Make a workspace and an agent
A workspace is the thing your agents share — their memory, your goals, the record. Sylvia asks what the agent is for and gives it a name, a face and a job. Give it a narrow job: "reviews changes before I do" beats "helps with engineering", because an agent with a narrow job can be told it got something wrong.
Copy your endpoint and token
Open Connect a client in the workspace. There is one endpoint for everything and one token per person — the token is who you are, so it is not shared between people.
Point a client at it
Claude Code takes it on the command line. Claude Desktop, Cursor, Zed, VS Code and Windsurf all take a block of JSON that looks the same in each. ChatGPT takes it as a connector in settings.
Talk to the agent by name
This is the part people skip. Address the agent, and the client loads that identity before it does anything — who it is, how your team works, what it has been corrected on. Ask a question without naming an agent and you get a model's guess; name one and you get something that already knows the answer.
What an agent can do
The whole surface.
Who am I, and how does this team work? Called first, every time, before reading a file or answering.
What do we already know about this? Searched before guessing.
Records a correction or a convention, so it holds next time.
Removes something that turned out to be wrong.
Brings memory in from another workspace, or from wherever you kept it before.
Finds the ticket the work belongs to, rather than asking you which one.
Creates a goal with its checks, plans the tickets under it, fixes a brief, marks a check met.
Starts and ends a run, and logs the steps, decisions and files in between. This is what writes the day.
Reads what the team has written up — specs, investigations, meeting notes, how-tos.
Files a new note where the team files things, not in a folder of its own invention.
Shows the procedures this agent is allowed to run.
Runs one, with its steps and its checks, on the record.
Finds what people said about work like this before.
Things worth knowing early
Three that save an afternoon.
The token is a person
Not a machine. Every write carries who did it, which is what makes the record worth reading — so give each person their own rather than sharing one.
Memory is plain text
Everything an agent knows is a sentence you can read, edit and switch off. If you cannot read it, you cannot trust it, and you certainly cannot take it with you.
It works without the app
The workspace is where you read the record and set goals, but an agent in a terminal with a token has everything it needs.
Ten minutes to the first one.
Make an agent, connect one client, and give it something small that you would otherwise have explained twice.