Most single-session chat workflows break down when you need specialized, parallel, or long-running work: one chat struggles to run isolated audits, keep background progress, or reliably hand off an implementation. This extension treats delegation as a first-class primitive for Pi: spawn focused child sessions that run a single job, return structured artifacts, and either stream results or keep working in the background.
What Sets It Apart
- Narrow child agents by design: each subagent gets a focused context and an explicit contract (scout, planner, worker, reviewer, oracle, etc.), which reduces prompt bloat and accidental orchestration drift. That makes reviews, research passes, and implementation steps easier to reason about and reproduce.
- Flexible execution modes: run agents in-foreground for streaming collaboration, run in-background for long tasks with async completion notifications, or fan-out parallel reviewers with worktree isolation to avoid git conflicts. The extension also writes per-run artifacts and session files for debuggability and revival.
- Practical orchestration features: supports model overrides per role, per-step config, saved chain/workflow files, optional intercom bridging for child→parent questions, and worktree hooks to preserve local build/test environments during parallel edits.
Who it's for & tradeoffs
Great fit if you orchestrate LLM-driven development workflows inside Pi and want repeatable delegation patterns — teams that need automated code review passes, parallel audits, or background implementation runs will get immediate value. It is less appropriate if you need a vendor-agnostic agent platform or a standalone multi-agent system outside Pi: it expects a Pi host environment, session persistence, and configured providers (so model usage and quotas remain a consideration). Nested delegation is intentionally limited by recursion guards to avoid runaway complexity.
Where it fits
Use it as an extension to add safe, auditable delegation to a Pi-driven workflow: think “ask an oracle for a second opinion,” “run parallel reviewers on this diff,” or “have a worker implement a plan and then run reviewers.” For standalone multi-agent orchestration across different chat hosts, consider purpose-built orchestration frameworks instead.
