Agents
What are agents
An agent is an AI delegate that does real work — it reads a repo, answers questions about the code, and (when you allow it) edits code and hands back a diff. You ask it right where the decisions are made, inside a topic. It runs either in the cloud — a sealed container with the repo already in it — or on its owner’s paired machine, and only ever within the repos and permissions set for it.
From a question to a change
Where the topic manager reasons about the work, an agent does it. Ask a question and it reads the repo and answers in the topic; give it a to-do in write mode and it makes the change in its own container or checkout and reports back with a diff you review — it never commits or pushes on its own.
The principles
- Scoped. An agent lives in one organization and can only reach the repositories allowed there — invisible everywhere else.
- Two homes. It runs in the cloud (a sealed container) or on its owner’s paired machine — same agent, different place the code lives. See Types.
- Permissioned. No silent work: a machine agent’s owner approves a new asker (or allows them always); a cloud agent runs for anyone in the topic, only in the repos the organization’s owner allowed.
- Set up in advance. Which repos it may touch, what it may do, and who may ask it are chosen when you create it, and can be changed later.
How you work with them
You reach an agent from inside a topic — @mention it by its @handle, or hand it a to-do. There are two kinds (see Types), both created under Agents & MCP — the laptop icon in the top bar. A cloud agent needs the organization’s GitHub connected; a machine agent needs a paired machine (see Agents on your machine).