AI Agents
What agents do in Elixion, and where to read about each part of them
Elixion is the platform for Engineering Intelligence™. Agents are the part of it that does work rather than tracks it: you assign an issue to an agent the same way you assign it to a colleague, and what comes back lands on the same board, in the same comment thread, and in the same delivery record.
This is the short version. Each section below links to the page that covers it properly.
Assign work to an agent
An issue's Assignee field has two groups, Users and AI Agents. Picking from the second records an agent assignee in its own field, so agent work is never counted as a person's.
Agent options need a repository connected to the project. Until then they are disabled and marked "(connect a repo)". Putting an agent on a project is still allowed without one, and the surfaces that do it say what is missing and link to the place to connect a repository.
See Creating Issues and GitHub.
Autonomy decides what happens on assignment
This is the single setting worth knowing before you assign anything.
- Fully autonomous. The run starts. This is the default.
- Acts with approval. The run is staged. The issue shows a banner naming the agent and a Start control, and nothing runs until someone presses it.
- On demand is carried by older agents and is not offered on the dropdown.
You set it from the agent's card on the Agents page. See AI Agents Overview.
What an agent produces
- Comments on the issue, in the same thread you use.
- A run record with status, duration, cost and tokens.
- Delivery evidence: pull requests with their merge state, commits, test runs with pass or fail, artifacts and links.
A merged pull request whose title or branch names an issue moves that issue to In Review and assigns it back to its reporter. Acceptance stays with a person.
Where to look
| Surface | What it is for |
|---------|----------------|
| Agents (/agents) | The roster: activity, runs in the last seven days, success rate, autonomy |
| Agent dossier | Performance, skills, workload, projects and assignment history for one agent |
| Agent log (/agents/log) | Every run in the workspace, filterable and exportable |
| Prompting (/prompting) | Drafting issues, sprints and documents from a sentence |
| Project Settings, Agents tab | Which agents can be assigned work in a project |
| Connect (/agents/connect) | Registering an agent you run yourself |
The roster
What you can assign is the roster on your Agents page: the built-in role agents, the agents you created, and the agents linked to projects you can reach. Agents you run yourself join through the Claude connector. See Available Agents.
Limits worth knowing
- An agent never picks up work by itself. Every run begins with an assignment, a schedule, or someone pressing Start.
- Through the Claude connector, agent tools are read only: list agents, list runs, read an agent's output. No dispatch, re run or reconfiguration.
- A success rate reading as a dash means no run finished in the window, not that everything failed. Utilisation reading as a dash means not measured, not zero.