QuintinShaw/pi-dynamic-workflows

What it solves

It addresses the limitations of single-agent context windows when performing broad, complex tasks—such as codebase-wide audits, large refactors, or deep research—that are too expansive for one prompt and one session to handle effectively.

How it works

Pi generates a deterministic JavaScript orchestration script that breaks a high-level request into smaller, manageable tasks. This script uses a set of runtime globals to manage the execution:

• Fan-out: It spawns multiple isolated subagents that can run concurrently (up to 16 at a time).
• Routing: Tasks are routed to different model tiers (small, medium, big) or specific providers based on the required complexity.
• Verification: The system uses built-in patterns like verify() and judgePanel() to cross-check results from multiple agents before synthesizing a final answer.
• Isolation: Agents can operate on separate git worktrees to prevent conflicts during parallel edits.
• Resumability: Work is journaled, allowing users to resume interrupted runs or replay only edited parts of a script without wasting tokens on unchanged steps.

Who it’s for

Developers and researchers using the Pi ecosystem who need to automate multi-step, multi-agent workflows for auditing, reviewing, or researching large volumes of data and code.

Highlights

• Parallel Orchestration: Ability to run up to 1,000 total subagents with high concurrency.
• Model Tiering: Dynamic routing between different model sizes to optimize cost and performance.
• Journaled Resume: Cache-aware replay that prevents redundant token spend after interruptions or script edits.
• Git Worktree Isolation: Safe parallel file editing via throwaway branches.
• Built-in Patterns: Pre-defined workflows for deep research, adversarial review, and codebase audits.
• Live Monitoring: A progress panel to track phases, token usage, and costs in real-time.