Repository navigation
07_Subagents
Imza can delegate work to concurrent sub-agents. Subagents have their own session and transcript, run inside the same working directory, and share your permission grants, operating under the same boundaries as the main agent.

| Role | Used for | Mode |
|---|---|---|
| Research | read-only discovery and code review | Plan |
| Builder | edits code and runs build tasks | Build |
| Basic | one-shot passes: session titles, AI Review, and Handoff | follows the task |
The main agent decides when to delegate; it can launch one to eight tasks in a single delegation, mixing research and build tasks. Build delegation is only allowed from a Build parent, and subagents cannot delegate further or manage todos themselves.
/subagents opens the role editor. Builder and Research take their own
model and reasoning effort; a role without an explicit model falls back to
the main selection. Basic covers the one-shot passes (session titles, AI
Review, and Handoff) and takes a model only: its passes are extraction
tasks with self-contained instructions, so it runs without reasoning by
default. Set an effort for it by hand in the configuration file if you
want one.
The editor writes the models section. See
Configuration for the
keys, the per-role defaults, and how reasoning effort is applied.
- Delegated tasks appear in the conversation with running, completed, or failed state.
- A subagent's Markdown transcript can be opened from its tool call in the chat even while it is still running.
- Completion reports are collected into the parent tool result, so the main agent summarizes them for you.
- Subagent approvals surface through the same modal queue as yours, labeled with the child agent's name.
- The status line shows how many agents are running.
Cancellation is cooperative: interrupting the turn stops subagents at safe boundaries.