The Know-How Graph
Agents are brilliant — but hopeless at repeatability
AI agents can research, plan, and code from scratch. They solve novel problems with remarkable creativity. But ask one to process a thousand expense reports the same way, and it'll improvise differently each time. The intelligence is there, but the consistency isn't.
The root issue: agents lack know-how memory. They retain facts — but they don't remember how they solved a problem previously, forcing costly reinvention every single run.
The gap in today's tooling
We've seen a wave of solutions try to tackle AI workflow repeatability, but each falls short in its own way:
- Code frameworks like LangChain or LlamaIndex let you build pipelines in Python — but they bury business logic in technical boilerplate. Every pipeline is custom, hard to maintain, and inaccessible to domain experts.
- Visual builders like Zapier, Make, or n8n are great for data shuffling, but they weren't designed for intellectual work that requires reasoning and understanding.
- Agent frameworks optimize for autonomy and exploration — exactly the opposite of what production repeatability demands.
None of these were designed for the core problem: formalizing how knowledge work gets done, in a way that humans, engineers, and AI agents can all understand and reuse.
A universal language for workflows
What's needed is a declarative language that expresses business intent rather than technical implementation — similar to how SQL separated what you want from how the database gets it.
Workflows written in this language should be:
- Consistent — same input produces the same output, every time
- Efficient — the right model is used for each step, not one-size-fits-all
- Transparent — auditable and self-documenting, not black boxes
- Composable — small methods combine into larger ones, like Lego bricks
Imagine a recruitment workflow that composes candidate profile extraction + job offer extraction + scoring logic — each piece reusable across different hiring pipelines.
The Know-How Graph
This is the vision: a network of composable, reusable workflow methods — executable procedures rather than static knowledge. Each node in the graph is a proven method. Each edge is a composition. Together, they form shared infrastructure for how organizations get work done with AI.
The Know-How Graph isn't just a library of prompts. It's a graph of executable know-how — methods that encode not just what to do, but how to do it reliably.
Organizations repeatedly solve identical problems in isolation. Standardized workflows could become shared infrastructure — much like open-source code libraries transformed software development.
Pipelex: an open-source implementation
Pipelex is our MIT-licensed, open-source implementation of this vision. Workflows are expressed in a declarative TOML-based syntax (.mthds files) and consist of modular pipes — knowledge transformers with guaranteed output structures.
The architecture is designed so that:
- Domain experts can read and validate workflow logic without writing code
- Engineers can version, test, and deploy methods like software
- AI agents can author, compose, and execute methods autonomously
The result: development time drops, operational costs shrink, and the gap between business intent and technical implementation narrows.
What this means
When proven methods become shared, composable artifacts — versioned in Git, executable by any agent, portable across organizations — we stop reinventing the wheel. We start building on each other's know-how.
That's the Know-How Graph. And it's open source.