Agents
An agent in Orbit Lumen is a saved helper with its own instructions, tools and projects, such as “write the release notes for our sprint”. This page is for members who build agents and for administrators who oversee them.
An agent reads only through the tools you allow and only in the projects you choose. It never changes anything by itself: a change is still a card that someone confirms. See Actions with confirmation cards.
Create an agent
Section titled “Create an agent”Open the Agents tab in the Lumen panel. An agent has these parts:
| Part | What it does |
|---|---|
| Name and description | How the agent is listed |
| Instructions | Added to the Lumen prompt after Lumen’s own rules |
| Allowed tools | The read tools the agent may use. Only tools of connected products are offered. |
| Projects | The Orbit projects the agent may read. Results from other projects are removed, and the agent cannot read or propose changes there. You can choose only projects you can access. |
| Visibility | Only you, the members of the agent’s projects, or everyone. Only an administrator can share with everyone, and a viewer cannot share. |
Pages of Folio spaces in other projects are removed from the agent’s results too.
Test and publish
Section titled “Test and publish”-
Save a draft.
-
Use Try in chat to ask the draft a question. Answers from a draft are not kept in your history.
-
Use Test questions. Add example questions with the words the answer should name, and run them on the draft. Lumen checks that each answer names what you expect.
-
Publish. The result of the test run is saved with the version.
Versions are numbered, and you can go back to an earlier one. Once an agent is published, the people it is visible to can ask it from the chat. As the author you see its answers and their ratings for 30 days. An administrator can see and delete any agent.
Built-in agents In development
Section titled “Built-in agents ”Lumen is getting seven built-in agents. You cannot change them, but you can copy one and adapt the copy. A built-in agent whose product is not connected is not shown.
| Agent | What it does |
|---|---|
| Work item triage | Looks at a new work item and checks for duplicates and urgency |
| Release notes | Writes release notes for a sprint |
| Code reviewer | Reviews a Forge merge request. It links the Orbit work item, lists problems by importance with file and line, gives a verdict and posts line comments as cards when you ask. It also checks the changes against team rules in .lumen/review.md, AGENTS.md or CONTRIBUTING.md in the repository. |
| Pipeline doctor | Reads the end of the log of each failed job in a failed pipeline, names the deciding error line, and says whether the cause is code, CI configuration or infrastructure. It proposes a fix as a card: a .gitlab-ci.yml change checked by CI lint, a merge request comment or an Orbit bug. |
| Incident analyst | Answers what failed, when it started and the likely cause, reading Pulse first and then Orbit, Forge and Folio |
| Doc writer | Drafts documentation from work items, for example a workaround found in a comment |
| Administrator helper | Helps administrators. Only administrators see it. |
Run an agent on events In development
Section titled “Run an agent on events ”An agent of yours, or a copy of a built-in one, can run when something happens:
- a new work item in Orbit;
- a new incident, or a triggered monitor, in Pulse;
- a new merge request, or a failed pipeline, in Forge.
You turn this on with Run on events in the agent editor. You choose the projects and give the agent access in Orbit. The run is read only: it uses a short-lived session limited to those projects, with no administrator tools. The same event never starts a run twice, and each agent has a daily limit on runs.
The answer and any cards arrive in your Inbox, which is a button with a counter in the panel. If a run cannot read something, the Inbox says why in words. Administrators can switch agent access off, which stops every agent session at once.