Subagents

A subagent is a focused specialist that carries out a well-scoped task — exploring the code, analyzing architecture, reviewing, implementing, testing — often in parallel and on the faster model. They let the main agent delegate clearly scoped work without losing the thread of the conversation.

Read-only subagents are spun up by the agent; the ones that write are launched by you via /agent spawn.

Two ways to run

• Automatic (the agent decides). For read-only work — exploring, analyzing, reviewing — the agent spins up a subagent on its own, synchronously. Only the explorer, analyzer, and reviewer types can be spun up this way.
• Manual (you launch it). In the Terminal (TUI), /agent spawn <type> <task> lets you launch any type — including the ones that write files (implementer, tester, refactorer, documenter).

Subagent types

Type	Role	Writes?	Typical tools
explorer	fast file/code search	no	read, search, list
analyzer	structure/architecture analysis	no	read, search, list
reviewer	review (security, performance, maintainability)	no	read, search
implementer	implement features	yes	read, create, edit, run
tester	create/run tests	yes	read, create, edit, run
refactorer	refactor code	yes	read, edit, run
documenter	documentation/READMEs	yes	read, create, edit

Per-model routing

Lightweight/read-only subagents (explore, analyze, review) run in a lighter, faster configuration — ideal for scanning and summarizing. Planning/implementation work inherits the session's model. See Models.

Commands (/agent, in the TUI)

Command	What it does
/agent or /agent list	Lists the session's subagents.
/agent spawn <type> <task>	Launches a subagent for a task.
/agent status [id]	State of one (or all) subagents.
/agent cancel <id>	Cancels a subagent.
/agent kill <id>	Force-terminates it.
/agent metrics	Counts, success rate, and tokens per type.
/agent cost	Token budget: total, consumed, and remaining.

Full reference in TUI commands.

Limits and budget

• Up to 8 concurrent subagents per session.
• A shared token budget across the subagents; once it's exceeded, new launches are refused.
• Limited depth: a subagent doesn't keep spawning others indefinitely.
• The automatic (read-only) subagent runs synchronously and has an output cap so it doesn't flood the context.

Use /agent metrics and /agent cost to track consumption and performance.

On each surface

• Terminal (TUI) — full control via /agent (plus the agent's automatic spin-ups).
• Desktop — the agent still uses read-only subagents automatically, but there is no /agent command for manual launching. (The Desktop task panel, Ctrl+T, is a task checklist, not the subagent system.)
• VS Code — Task Mode is a separate feature: a plan → step → report loop over your normal chat session, not Imaginne's local subagent system. See Using in VS Code.

See also

🧠Models

Why lightweight subagents use a lighter configuration.

Open →⌨️TUI commands

The full list of /agent and its subcommands.

Open →🧩Extending Imaginne

The other extension points.

Open →