Roadmap

​

Current status

Where pflow is today (v0.8.1):

• Core workflow engine built on PocketFlow
• Markdown workflows — .pflow.md files that agents read and write naturally
• Node system — file, llm, http, shell, code (Python), claude-code, and MCP bridge
• AI agent integration via CLI and MCP server
• Discovery — find nodes and workflows by describing what you need
• Template variables — ${var} syntax connecting node outputs to inputs
• Workflow validation with error messages designed for agents, not humans
• Batch processing — process arrays through nodes, sequential or parallel
• Unix-first piping — stdin/stdout, works with any Unix tool
• Skills publishing — save workflows as Claude Code skills, cross-platform
• Execution traces for debugging
• Settings management — API keys, node filtering
• Unified model support — any provider via Simon Willison’s llm library
• Published on PyPI — uv tool install pflow-cli

​

Now

Workflow expressiveness Right now workflows run top-to-bottom, every step in order. That’s fine for simple pipelines but it breaks down fast — you can’t skip a notification step when there’s nothing to notify about, you can’t reuse a workflow you already saved inside a bigger one, and you definitely can’t let a workflow send a Slack message or create a PR without someone approving it first.

• Conditional branching — if/else logic so workflows can take different paths based on data
• Nested workflows — saved workflows become nodes inside other workflows. The changelog workflow you built last week becomes one step in your release workflow.
• Human-in-the-loop approval gates — pause execution for human review before continuing. Without this, workflows that take real-world actions aren’t trustworthy enough to actually use.
• Smart trace debugging — when a workflow fails, the system knows what each node was trying to do because node types are finite. “HTTP node expected 200, got 403” or “template referenced a field that doesn’t exist” — not a raw JSON trace the agent has to parse.

​

Next

Iteration speed and output quality Building workflows means iteration — run it, see what breaks, fix it, run again. Two things make that loop slower than it needs to be.

• Iteration caching — when an agent fixes step 7 and re-runs, steps 1-6 shouldn’t re-execute. Node configs get hashed, outputs cached, cache invalidated when upstream changes. This only works because workflows are made of discrete nodes — you can’t hash and cache parts of a script.
• Structured output — LLM and Claude Code nodes return typed JSON instead of free-form text. Downstream nodes stop guessing about the shape of what they received.
• Sandbox runtime — isolated execution for shell and code nodes. Needed before running anything an agent generated that you haven’t reviewed.

​

Later

• Code export — export a workflow to standalone Python, zero pflow dependency. Build and iterate with structure, ship plain code.
• Workflow testing — mock nodes, assert outputs, pflow test. Right now modifying a saved workflow and re-publishing it is a trust exercise.
• Task parallelism — fan-out/fan-in for independent nodes. Different from batch processing, which runs the same node on multiple items.
• Execution preview — show what will run, in what order, with what data flow, before anything executes.

​

Vision

pflow is infrastructure, not a destination. It provides building blocks and a runtime — agents do the assembly. The better the building blocks get, the more capable the agents become. The longer-term direction:

• Workflows as distributable packages — export as standalone MCP servers that work without pflow installed
• Dynamic MCP discovery — search and install MCP servers on demand instead of manual configuration
• Progressive disclosure — the agent instruction file is ~1900 lines today. Most of that should live in the nodes themselves and load only when needed.

pflow improves through a direct loop: agents build workflows, I find where they struggle, I fix it. Because the surface area is finite — node types, template syntax, error messages — each fix is targeted and compounds. A year from now pflow will be meaningfully better at helping agents build workflows, not because of some grand vision but because each friction point gets filed down one at a time.

​

Get involved

---

Built by a developer who got tired of watching agents re-think the same tasks.
Questions or ideas? Reach out — andreas@pflow.run