Sub-agents & orchestration

Pi ships without sub-agents by design — but they’re one of the most popular things people add back. Sub-agents run a task in an isolated context window and report a summary, which keeps your main session focused and saves tokens. A few mature options exist, from lightweight delegation to full orchestration.

Alternatives

1. pi-subagents (nicobailon)

The most-installed sub-agent extension. Delegates tasks to sub-agents with chains, parallel execution, and a TUI clarification flow.

pi install npm:pi-subagents

Benefits

• By far the most popular option (highest downloads in this category).
• Parallel and chained execution out of the box.
• Interactive clarification UI when a sub-agent needs input.

Drawbacks

• More features means more surface area / config to learn.
• Parallel sub-agents multiply token spend — watch your budget.

2. @tintinweb/pi-subagents / @gotgenes/pi-subagents

Claude Code-style autonomous sub-agents. @gotgenes/pi-subagents is a “friendly fork” of @tintinweb/pi-subagents with additional polish; both aim to replicate the Claude Code sub-agent experience.

pi install npm:@tintinweb/pi-subagents
# or the fork:
pi install npm:@gotgenes/pi-subagents

Benefits

• Familiar mental model if you’ve used Claude Code sub-agents.
• The fork tracks fixes/features on top of the original.

Drawbacks

• Two similar packages — pick one to avoid duplicate tools.
• “Autonomous” agents need clear task scoping to avoid wandering.

3. @pi-agents/orchid (orchestration)

When you’ve outgrown one-off delegation: a unified orchestration package combining a batch orchestrator, a “Ralph loop”, and subagent dispatch for multi-agent pipelines.

pi install npm:@pi-agents/orchid

Benefits

• Built for pipelines, not just single delegations.
• Batch + loop patterns for repetitive multi-step work.

Drawbacks

• Heavier and more opinionated — overkill for simple “go research X” tasks.
• Steeper learning curve.

Which should I pick?

• Most people: start with pi-subagents — popular, parallel, well-rounded.
• Coming from Claude Code: @gotgenes/pi-subagents (or the upstream @tintinweb version) feels familiar.
• Building multi-agent pipelines: graduate to @pi-agents/orchid.

Tip

Sub-agents shine for bounded research (“analyze these three modules”) and parallelizable work. For a single linear task, the main agent is usually faster and cheaper.