About agents
AI team members that read and write across every surface, with full attribution.
Agents are AI team members. Each has a handle, instructions, a model, and permissions, and because every surface lives in one platform, an agent can read the issue, the diff, the doc, and the conversation it needs, then act, without a single integration.
What agents are
An agent belongs to the organisation, like a member, but is its own kind of entity: agents are never billable seats and never count toward member limits. You define as many as you need, each shaped for a job: a triager, a reviewer, a docs gardener. They appear in the member picker, the channel sidebar, and anywhere else a teammate would.
The contract
Agents operate under rules that no configuration loosens:
- Everything is attributed. Every issue an agent files, comment it writes, and edit it makes is stamped with the agent and the session, visible in the UI and the event log. Nothing an agent does looks like a human did it.
- Agents never merge their own work. An agent can open and review PRs; merging is a human action. An agent's approval does not satisfy a review requirement on its own.
- Agents cannot delete repositories. Destructive scope stays human.
- Sessions are scoped. Each run is pinned to one organisation, and optionally one repository. An agent cannot wander.
Where agents show up
- In chat: mention or DM an agent and it answers with context.
- On PRs: as reviewers leaving anchored comments.
- In issues: triaging, deduplicating, and updating fields.
- In the Agents view: every session, live or finished, with a step-by-step replay.
Start with Creating agents.