Agents
Natamux starts five coding agents with hooks, and runs any other terminal agent without them. A task has one agent. A stage can start another, with its own model. See Mix agents in one task.
Compatibility
Section titled “Compatibility”| Agent | Program | How it reports | Failed state | Skills by name | Tested |
|---|---|---|---|---|---|
| Claude Code | claude |
HTTP hooks, by --settings |
Yes | /name |
Real sessions |
| Codex | codex |
-c hooks.<Event>=… overrides |
No | $name |
Not run against a real session |
| Gemini CLI | gemini |
A settings file, by environment variable | No | Path of a SKILL.md |
Not run against a real session |
| Copilot CLI | copilot |
.github/hooks/natamux.json in the worktree |
Yes | /name |
A real session, 1.0.94 |
| Cursor CLI | cursor-agent |
.cursor/hooks.json in the worktree |
Yes | Path of a SKILL.md |
Not run against a real session |
| Other agents | any | The terminal bell, natamux ask and natamux done |
No | Path of a SKILL.md |
Not applicable |
- Every agent that has hooks reports when it works, asks and finishes. The hooks are for that session only. Your own settings do not change.
- The Cursor CLI has no hook for a wait on permission, so the terminal bell marks that wait.
- “Not run against a real session” means the hooks follow the vendor’s documentation and are tested with made-up input.
- Each page has the exact command line, the events and the known limits.
Which agent for which stage
Section titled “Which agent for which stage”Natamux does not rank agents. Pick by what the stage needs:
| Stage | What helps |
|---|---|
| Implement | The agent you trust to change code. It runs in Shell 1. |
| Review, Security review | A different agent than the one that wrote the code, so the review has fresh eyes. These stages run in their own shell. |
| Test spec | An agent that does not read the code. It runs beside Implement. |
| A CI stage of your own | Any agent that can read a failed log. |
Set agent and args on the stage in .natamux.json.