AgwaB/pi-workflow

What it solves

pi-workflow provides a way to run complex, multi-step AI agent workflows that are repeatable and structured. Instead of relying on a single prompt, it allows users to define a deterministic process—such as deep research or code review—where multiple sub-agents can work in parallel or sequence to achieve a high-quality result.

How it works

The system organizes work into a "stage graph" where a natural-language task is processed through a series of defined steps. It uses @agwab/pi-subagent to coordinate workers and supports several execution patterns:

• Single: A single focused prompt for one sub-agent.
• Foreach: A dynamic fan-out where one agent creates a list of items, and separate agents process each item.
• Reduce: A fan-in synthesis where results from multiple upstream steps are combined into a final report.
• Loop: Bounded repetition until a stop condition is met.
• DAG: Nested graphs for complex organizational structures.
• Dynamic: Adaptive orchestration using controller code to create tasks on the fly.

Who it’s for

Developers and teams using the Pi ecosystem who need to automate recurring, high-stakes tasks like architecture reviews, spec conformance checks, and project-specific team routines.

Highlights

• Bundled Workflows: Includes pre-defined templates for deep-research, deep-review, spec-review, and impact-review.
• Visual Monitoring: A built-in TUI workflow board (/workflow) to inspect run progress, drill into stages, and view task artifacts.
• Customizable: Users can create or adapt workflows using the workflow-guide skill.
• Execution Profiles: Supports custom-named profiles (e.g., low, medium, high) to control how a workflow is executed.
• Adaptive Routing: An execution-router skill helps decide whether a task requires a single agent, a specific workflow, or a targeted verifier.