Pi Subagents 🤖 Agent Open source

What it does

pi-subagents is an extension for the pi coding agent that adds Claude Code-style subagents and scripted orchestration. The main agent gets an Agent tool that spawns specialised agents in isolated sessions, each with its own system prompt, model, thinking level and tool set. Agents run in the background by default and report back when they finish, and you can watch, steer or resume them from the terminal interface. Tool names and calling conventions deliberately mirror Claude Code, so prompts and habits transfer easily.

What is inside

• Tools: Agent for spawning or scheduling agents, get_subagent_result, steer_subagent for mid-run instructions, and SubagentWorkflow, which runs a JavaScript script with agent(), parallel() and pipeline() helpers for deterministic multi-agent flows
• Custom agents: Markdown files with YAML frontmatter in .pi/agents/ or .agents/agents/, with model choice, tool allow and deny lists, skill preloading and optional nested subagents
• Interface: a live widget, a FleetView list of running agents, a conversation viewer and @agent mentions that send a message straight to a subagent
• Isolation and memory: optional git worktrees per agent with automatic branch commits, and persistent memory at project, local or user scope
• Control: concurrency limits, graceful turn limits, model-scope enforcement, cron-style scheduling, and events plus RPC so other pi extensions can drive subagents

Works with

It is built and tested for pi only. pi is not one of the apps listed in this directory, and the README mentions a third-party adapter for another harness that the author does not test.

How to install

pi install npm:@tintinweb/pi-subagents

It requires a recent pi release, which the package declares as a peer dependency.

Maintenance and safety

The extension is MIT-licensed and actively maintained, with detailed documentation of every setting. Background agents run with the tools you allow, so restrict tools for untrusted tasks and review worktree branches before merging. Parallel agents multiply token use; the concurrency limit helps keep that in check.

Who should use it

pi users who want to delegate exploration, review or implementation to parallel background agents, or script repeatable multi-agent pipelines. People who do not use pi have no use for it.