Sub-agents

Spawn specialized sub-agents with chains, parallel execution, TUI clarification, and async support.

What it does

Sub-agents lets Pi spawn isolated child agents that each run in their own context. You can run agents in sequence (chaining), in parallel, or asynchronously — the orchestrator waits on results and assembles them however you specify.

Each agent gets a focused system prompt and a specific task. This keeps individual context windows small and targeted, which tends to produce better results than asking a single agent to hold too much state at once.

Pi-subagents ships seven built-in specialists: scout (codebase exploration), planner (task decomposition), worker (implementation), reviewer (code review), context-builder (context gathering), researcher (information lookup), and delegate (general-purpose forwarding). You can override any of them or add your own via agent frontmatter.

Why it's included

Complex tasks benefit from decomposition. A single monolithic agent working through a large refactor will lose track of things that a coordinated set of specialists handles cleanly. Sub-agents makes Pi's architecture match the way hard problems are actually structured — break it down, parallelize what you can, and review the result.

Commands

The primary interface is the subagent tool, which Pi invokes directly. Most interaction happens naturally through conversation rather than explicit slash commands.