The PromptFlow™ System is a persistent re-usable process layer that evolves along with the underlying models in a federated remunerative ecosystem. It sits above MCP servers, n8n workflows, LangChain agents, hosted models, and custom pipelines — providing what they individually lack.
Every AI framework claims to solve the same problems. The PromptFlow™ System doesn't compete with them — it measures them. It's the infrastructure above the substrate that makes federation actually work.
Every execution produces evidence — completion rate, cost, latency, schema conformance, audit verdicts. The graph doesn't trust claims; it trusts measurement.
Once every capability has a typed interface and measured behavior, building a new one becomes picking the right nodes and wiring them — not writing glue code.
When a new model can do in one call what previously took five steps, the older path becomes alternate. Complexity sheds automatically. This mechanism is unique to the PromptFlow™ System.
Every execution pays every party whose work contributed — the composer, the sub-process authors, the MCP operators, the eval contributors, the auditors. Proportional to contribution.
The graph is agnostic about implementation. MCP servers, n8n workflows, agents, hosted models, `.promptflow` files, and typed HTTP services all become addressable capabilities with the same interface, the same verification, and the same settlement treatment.
Tools surfaced via tools/list, registered as capability nodes. Every operator gains discovery and revenue.
Workflows with declared I/O schemas become composable capabilities. Creators gain distribution and monetization.
LangChain, AutoGen, CrewAI, or custom — any agent with an addressable endpoint. Framework choice is invisible to the graph.
Anthropic, OpenAI, Google, open-source — every model API is a provider type. Calls pass through at cost; settlement is metered.
Authored files that compose any of the above. The universal contributor path — anyone can build one without infrastructure.
Generic endpoints with request/response schemas. If it's accessible and typed, it composes.
Six load-bearing mechanisms validated end-to-end in a reproducible simulation with real LLM decisions: distinctness classification, autonomous graph growth via composition, ranking convergence, spot flips, simplification when newer implementations dominate older ones, and multi-party settlement with lineage royalties.
The specification is open. The reference implementations are MIT. The trademark and system are protected. The graph is meant to be owned by the people who build it.