Moirai Flow
Deterministic agentic workflow engine
My founding project: a deterministic, governance-first workflow engine for AI agents — turning agent pipelines into reproducible, auditable executions instead of probabilistic black boxes. Built for safety- and compliance-critical teams where unpredictable agent behaviour is a non-starter.
Problem
Agent frameworks optimise for autonomy: give the model tools, let it decide. That's fine for demos and fatal for safety- and compliance-critical organisations, where the same input producing a different execution path on every run means you can't audit, can't reproduce incidents, and can't certify anything.
Teams in those environments end up rejecting agents entirely — not because the models aren't capable, but because probabilistic orchestration is un-governable by design.
Approach
Moirai Flow inverts the default: workflows are compiled, not improvised. A pipeline is declared up front, the engine compiles it into a deterministic execution plan, and LLM calls happen inside fixed, observable steps — the model fills in content, it doesn't choose the control flow.
Governance is first-class rather than bolted on: strict rules and human gates are part of the pipeline definition, and every execution is reproducible and auditable after the fact.
Built on a TypeScript/Node core with MCP for tool access and provider-agnostic multi-LLM support, drawing directly on the runtime patterns I built for Medusa at AIBC.
Outcomes
- Working prototype running on an existing agentic stack.
- Currently in stealth while the engine and first target use-cases take shape.