Editing an agent
Name, prompt, key, disabling and deletion
Name
Edited straight in the heading of the agent's page: click it. The name has to be unique: if a taken name goes to the server at creation, the agent is created under a slightly changed one and the wizard says so.
Description and instructions
The General tab. The description is for you and for the agent list. The instructions are the system prompt: how the agent should behave. They are edited freely and at any time, and the new version applies from the next message.
Instructions can be cleared: save an empty field and the prompt goes back to "not set". The name cannot be cleared.
The agent key
The key is shown once, at creation, and applies only to external agents, the ones whose brain
lives on your side: CENTRIFUGO, WEBHOOK and MCP. For them it opens calls to the
platform: an MCP client authorises with the header Authorization: Bearer <key>, and a
websocket or webhook runtime carries the same one.
For a GENERIC agent that door is shut: its brain is the platform's own worker, and there is
no legitimate caller reaching the API on its behalf from outside. The dashboard chat needs the
key even less.
GENERIC does have one use for it, and it runs the other way: an IDE connects to the agent
over ACP with that key. The Coder role is built on this.
"Regenerate Key" issues a new one and immediately invalidates the previous one. Everything that travelled with the old key stops working that same second, so update your clients first.
Disable
The "Enabled" switch on the agent's page. A disabled agent does not work: its key stops letting anyone in and its runs do not execute. Settings, skills, connections and the journal stay where they are; switch it back on and everything continues.
This is what to do instead of deleting, when an agent is simply in the way.
Delete
Deleting an agent removes its settings and its bindings. The skills and connections themselves remain: they are yours, the agent was only using them.