A developer named Michaelliv has published pi-dynamic-workflows, a Pi extension that ports Claude Code’s Workflow tool to the open-source Pi coding-agent harness. The extension ships a workflow primitive that lets an assistant write a small JavaScript script, fan work out across dozens of isolated subagents, and synthesize the results. The capability was previously accessible only inside Anthropic’s own Claude Code CLI.
Pi is Anthropic’s open-source coding-agent harness, separate from the commercial Claude Code product. The pi-dynamic-workflows extension adds a single tool primitive: the assistant emits a JS script declaring parallel agent calls, the runtime executes the script deterministically, and the results come back in a form the orchestrating agent can synthesize. The script shape is compact, using export const meta = {...} followed by a body describing how to distribute work.
Subagents in a fanned-out workflow are not stripped-down runners. Each one gets file read, shell command, and structured-output tool calls, which means it behaves like a normal Pi conversation turn. That capability set covers the core use cases Michaelliv documents in the README: codebase audits where one agent handles each directory, multi-perspective code review using a find-then-verify-then-synthesize chain, large refactors with worktree isolation per file, and fan-out research where one agent processes each source.
The runtime enforces two hard limits. Concurrency is capped at min(16, CPU cores minus 2), which prevents a single workflow from saturating a developer machine. Total agent count per workflow lifetime is capped at 1000. Workflows are journaled: if a run crashes mid-execution, the resume feature re-runs only the changed or new agent calls rather than starting over. That last detail matters for long-running audits, where a crash at step 80 of 200 would otherwise waste significant compute.
The project is explicitly described in the GitHub repository as a derivative of Claude Code’s Workflow tool. That lineage is worth noting. Anthropic introduced the Workflow primitive inside recent claude-code-vNN releases as a way to orchestrate parallel agent work without building a custom orchestration layer from scratch. Michaelliv’s contribution is essentially a ports job: the same design philosophy, running on an open stack that teams can inspect, fork, and modify without a commercial license.
The pattern itself, a declarative JS workflow that fans out to deterministic agent execution and synthesizes back to a single result, is converging across agentic products. Cognition’s Devin and OpenAI’s Codex both organize parallel agent work around similar fan-out structures. The fact that this shape is now replicable in open-source at roughly the same capability level as the proprietary version is a signal that the primitive has stabilized.
For teams running Pi as their coding-agent harness, pi-dynamic-workflows offers Claude-Code-style fan-out without committing to Anthropic’s commercial CLI. The tradeoffs are real: the extension is a community project without official Anthropic support, and the resume capability lacks the telemetry and observability that the official Workflow tool provides. Teams running regulated or audited development environments should weigh that gap before adopting it for production workflows.
For teams already evaluating multi-agent orchestration for codebase audits or parallel code review at scale, pi-dynamic-workflows is worth benchmarking against the proprietary alternative. If the concurrency cap and agent limit fit your workload, the open-source path removes one vendor dependency from a stack that currently has several.
Source: GitHub repository pi-dynamic-workflows by Michaelliv, published approximately May 29, 2026.
