Subagents

How a child runs

During a turn the model calls the subagent tool with { agent, prompt }. lugent starts a child session, in the same store and process, with the subagent's own prompt, model and tools. The child's final message comes back as the tool result.

---
description: Probes one file and summarizes it
mode: subagent
tools: [read]
---
You are a probe. Read exactly one file and summarize it in three lines.

Rules

Discovery

A # Subagents section of the prompt lists id and description, and only when at least one is launchable.

Fresh context

The child shares no history, so the prompt must be self-contained.

Parallel

Several subagent calls in one tool round run at the same time.

Depth 1

A child never gets the subagent tool or the catalog, so it cannot launch a grandchild.

Hidden

Children are persisted with a parent link but stay out of /resume and the picker.

Not in /agent

Subagents are launch targets, not session personas.

The tool defaults to ask, and an agent's tools or permissions can allow, ask or deny subagent by name like any other tool.