Run any Skill in Manus with one click

Get Started

sub-agents

Stars2

Forks0

UpdatedMarch 18, 2026 at 12:50

Use multiple pi sub-agents for delegation, isolation, and parallel work. Use when tasks can be split and later merged.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

szokeasaurusrex

szokeasaurusrex/pi-agent

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

File Explorer

2 files

SKILL.md

readonly

name	sub-agents
description	Use multiple pi sub-agents for delegation, isolation, and parallel work. Use when tasks can be split and later merged.

Sub-Agents

Use sub-agents for focused delegation, parallel work, role separation, prompt isolation, and staged pipelines.

Uses

Single delegation for a focused subtask
Parallel delegation for independent subtasks
Role-based agents (researcher, implementer, reviewer)
Pipeline orchestration (research → implement → verify)
Fan-out/fan-in (split work, collect outputs, synthesize)
Model/tool specialization per sub-agent
Prompt/session isolation to reduce cross-contamination
Reliability controls (timeouts, retries, explicit output formats)
Advanced steering/cancellation via RPC mode

Launch

Use scripts/run-subagent.sh.
Use non-interactive mode only.
Pass --prompt "<prompt>" on every launch and resume.
Record and report the session path.
Inspect wrapper behavior with scripts/run-subagent.sh --help.
Inspect pi CLI help with scripts/run-subagent.sh -- --help.

Prompt

Make the prompt specific. State the exact task, scope, inputs, constraints, expected output, and stopping condition.
Avoid broad prompts such as look into this or help with this.

Thinking

The script defaults to medium. This thinking level is suitable for most tasks.
Set --thinking minimal or --thinking low for mechanical edits, straightforward execution, or format conversion, as less thinking is faster.

Recovery

If a sub-agent fails or times out, continue the session.
- Note: thinking level does not persist on the new call.
You may add a new prompt when continuing the session. Here are some useful sample prompts you can use with some tweaking
- If you just want to let the subagent continue, say, "You were interrupted, please continue" (useful for complex tasks).
- You can find out where the agent is by asking, "You were interrupted. Give a quick status update. Do not proceed until I ask you" (with minimal/low thinking level), then, based on the reply, decide what to do next.
- You can speed things along by saying, "You were interrupted. Stop researching/planning/etc and state what you know given your current knowledge and what open questions remain" (again, a lower thinking level is suitable).
If the subagent produces nothing helpful, you may consider starting a new subagent incorporating learnings from the previous one, or you can finish the task yourself.

Invocation

Launch: scripts/run-subagent.sh --prompt "<prompt>"
Launch with explicit thinking: scripts/run-subagent.sh --thinking minimal --prompt "<prompt>"
Resume: scripts/run-subagent.sh --resume ~/.pi/agent/subagent-sessions/<file>.json --prompt "<prompt>"
Extra pi args: scripts/run-subagent.sh --prompt "<prompt>" -- --model sonnet
Pi help: scripts/run-subagent.sh -- --help

More from this repository

same repository

delegation-plan

szokeasaurusrex/pi-agent

Write plans intended for a cheaper or less capable implementation model. Use when the user asks for a plan suitable for another LLM, junior engineer, intern, cheaper model, coding agent, or implementation delegate.

2026-07-062

thinking-through

szokeasaurusrex/pi-agent

Use when the user asks for ideas, brainstorming, discussion, exploration, possibilities, pros and cons, tradeoffs, or to "think through" something.

2026-07-032

plan

szokeasaurusrex/pi-agent

Use this skill when asked to plan something, or plan and implement. Do not use when asked to implement a plan.

2026-05-112

cargo

szokeasaurusrex/pi-agent

Use when initializing Rust projects or modifying Cargo.toml, Cargo.lock, cargo config, or other cargo-related files. Prefer this skill for dependency changes, features, package metadata, workspaces, and cargo commands. Do not load it for editing Rust source code generally.

2026-05-112

split-commit

szokeasaurusrex/pi-agent

Split an existing commit or PR into smaller reviewable pieces through a one-hunk patch queue. Use when extracting part of a commit, splitting a commit into multiple commits, or splitting PR changes from one or more source commits without starting from direct source edits.

2026-04-152

context-fork

szokeasaurusrex/pi-agent

Checkpoint, inspect, and return through session tree branches. Use when exploring uncertain approaches, investigating code, or when the user asks for a forked agent.

2026-03-192

Sub-Agents

Use sub-agents for focused delegation, parallel work, role separation, prompt isolation, and staged pipelines.

Uses

Single delegation for a focused subtask

Parallel delegation for independent subtasks

Role-based agents (researcher, implementer, reviewer)

Pipeline orchestration (research → implement → verify)

Fan-out/fan-in (split work, collect outputs, synthesize)

Model/tool specialization per sub-agent

Prompt/session isolation to reduce cross-contamination

Reliability controls (timeouts, retries, explicit output formats)

Advanced steering/cancellation via RPC mode

Launch

Use scripts/run-subagent.sh.

Use non-interactive mode only.

Pass --prompt "<prompt>" on every launch and resume.

Record and report the session path.

Inspect wrapper behavior with scripts/run-subagent.sh --help.

Inspect pi CLI help with scripts/run-subagent.sh -- --help.

Prompt

Make the prompt specific. State the exact task, scope, inputs, constraints, expected output, and stopping condition.

Avoid broad prompts such as look into this or help with this.

Thinking

The script defaults to medium. This thinking level is suitable for most tasks.

Set --thinking minimal or --thinking low for mechanical edits, straightforward execution, or format conversion, as less thinking is faster.

Recovery

If a sub-agent fails or times out, continue the session.

Note: thinking level does not persist on the new call.

You may add a new prompt when continuing the session. Here are some useful sample prompts you can use with some tweaking

If you just want to let the subagent continue, say, "You were interrupted, please continue" (useful for complex tasks).
You can find out where the agent is by asking, "You were interrupted. Give a quick status update. Do not proceed until I ask you" (with minimal/low thinking level), then, based on the reply, decide what to do next.
You can speed things along by saying, "You were interrupted. Stop researching/planning/etc and state what you know given your current knowledge and what open questions remain" (again, a lower thinking level is suitable).

If the subagent produces nothing helpful, you may consider starting a new subagent incorporating learnings from the previous one, or you can finish the task yourself.

Invocation

Launch: scripts/run-subagent.sh --prompt "<prompt>"

Launch with explicit thinking: scripts/run-subagent.sh --thinking minimal --prompt "<prompt>"

Resume: scripts/run-subagent.sh --resume ~/.pi/agent/subagent-sessions/<file>.json --prompt "<prompt>"

Extra pi args: scripts/run-subagent.sh --prompt "<prompt>" -- --model sonnet

Pi help: scripts/run-subagent.sh -- --help