Pi Dynamic Workflows 🤖 Agent Open source

What it does

pi-dynamic-workflows, by Quintin Shaw, brings script-driven multi-agent workflows to the Pi coding agent. When you ask for a workflow, Pi writes a short JavaScript orchestration script that fans work out to fresh subagent sessions, routes each one to a suitable model, cross-checks the results and returns one answer. Intermediate results live in script variables rather than your chat context, which suits codebase-wide audits, large refactors and multi-perspective reviews.

What is inside

Scripts use a small set of runtime functions: agent(), parallel(), pipeline(), phase() and workflow() for calling saved workflows. Quality helpers include verify(), judgePanel(), loopUntilDry() and completenessCheck(). Each agent call can choose a model tier or an exact model, a thinking level, a JSON schema for structured output, and git worktree isolation so parallel edits land on separate branches. Completed agent calls are journaled, so an interrupted run can resume without paying for finished work again, even after you edit the script. A live panel shows phases, models, tokens and cost, and runs can be saved as reusable commands. A packaged workflow-authoring skill documents the rules.

Works with

This is a Pi extension. Pi is not one of the apps in this directory, so no listed app can install it directly, although the design mirrors Claude Code's dynamic workflows.

How to install

pi install npm:@quintinshaw/pi-dynamic-workflows

Maintenance and safety

MIT licensed, published on npm and pushed within the last week, with upgrade notes between major versions. Workflow mode is armed by the word "workflow" in a message by default; you can change or disable that trigger. Large fan-outs can spend many tokens quickly, so set run or phase budgets when you need a ceiling.

Who should use it

Pi users who regularly run broad audits or refactors that exceed one context window and want deterministic, resumable orchestration. Casual users can stick with single subagents.