Hi everyone,
I’ve been building BoundFlow, an open-source operational control plane designed to sit on top of frameworks like LangChain and LangGraph to help run AI workflows safely in production.
The idea isn’t to replace existing agent frameworks, but to separate business logic from operational concerns.
For example, instead of embedding operational logic directly into your workflow, BoundFlow lets you define things like:
- Runtime policies (tool/model restrictions, cost/token budgets, approvals, etc.)
- Lifecycle policies (automatically pause, roll back, or disable unhealthy workflows)
- Durable workflow execution
- Operational audit logs
- Versioning and rollouts
- Customer-owned execution (your code continues running in your own infrastructure)
The control plane coordinates execution while customer-defined operations continue running in their own workers.
I’m curious what people think about this architectural direction.
Project: https://boundflow.dev
Examples using LangGraph/LangChain agents: **https://boundflow.dev/docs/integrations/
**
I’d really appreciate any architectural feedback or criticism.