Sub-agents

The agent can hand a self-contained task to a sub-agent that works in its own context and returns a short report, so the main conversation stays small. explore is built in: a read-only researcher that searches the code or the web and cites path:line. It runs on the fast model when the host has one and on the main model otherwise. The host picks the fast model in the settings (F2 → Model): auto takes a cheaper model you can already reach, off keeps everything on the main model, or choose one; PLURAL_FAST_MODEL does the same from the environment. Several sub-agents can run in parallel.

A sub-agent shows up as one Agent line in the conversation with what it is doing right now; ctrl+o opens everything it looked at, and the line ends with the model it used, its tool calls and its tokens.

Creating your ownCopy Anchor Link

/agents lists them and + create an agent… asks for a name, when to use it, its instructions, what it may touch (read only, read and run commands, or read, edit and run), which model and where to save it. Or add your own as Markdown files in .agents/agents, .claude/agents or .plural/agents, in the project or the home folder. The format is the one Claude Code uses:

---
name: reviewer
description: Reviews a diff for bugs and missing tests.
tools: Read, Grep, Glob, Bash
model: haiku        # haiku/fast runs on the fast model, anything else on the main one
max_turns: 20
---
You review the diff you are given for another agent.

<task>
- Look for bugs, missing tests and changes the diff does not explain.
</task>

<report>
- One line per problem, with path:line.
</report>

Plain instructions work too. Parts in XML tags, as Plural's own prompts and the built-in explore (src/plural/agent/builtin_agents/explore.md) use, help the model tell the parts apart.