BOLT Agentic Payments Control Plane

Inspiration

AI agents are increasingly able to recommend and initiate consequential actions. Payments are a clear example where useful automation needs an independent control layer: model reasoning should not automatically become financial authority.

What it does

BOLT Agentic Payments Control Plane is a governed execution layer for AI-assisted PayPal workflows.

The prototype accepts a structured payment or refund proposal and evaluates it against deterministic policy before anything can execute:

• low-risk actions can be allowed within a configured autonomous limit;
• consequential actions return approval_required and pause for a human;
• actions outside policy are denied;
• supported PayPal Sandbox actions execute only when valid sandbox credentials are explicitly configured.

This creates a clear boundary between AI intent and payment authority.

How we built it

The current prototype is a small open-source Node.js service with an explicit policy engine, spending and currency limits, allow / approval-required / deny decisions, a PayPal Sandbox OAuth and Orders API integration path, environment-based credential handling, and automated tests for the policy layer.

The next build step is the AI planner: natural-language intent will be converted into a structured proposed PayPal action. BOLT will remain the independent policy and approval layer between that AI output and PayPal.

Challenges

The core challenge is balancing useful agent autonomy with financial control. A payment agent needs more than a good prompt: it needs hard limits, explicit approval points, predictable failure behavior and a traceable record of what it attempted to do.

We deliberately keep credentials out of the model path and separate policy decisions from generative reasoning.

What we learned

For agentic payments, the important architecture is: AI intent -> structured action -> deterministic policy -> human approval when required -> PayPal Sandbox execution -> audit trail.

What's next

Before final submission we are completing the natural-language AI intent-to-action planner, a human approval interface, PayPal Sandbox end-to-end order/capture and refund flows, persistent audit events, a hosted functional demo, and a demonstration video under three minutes.

This is a hackathon prototype for PayPal Sandbox and representative test data. It is not a production payment processor or certified compliance product.

Source

Public MIT-licensed repository: https://github.com/ondmindmanagement-hub/bolt-paypal-control-plane

Built With

• agents
• ai
• fintech
• javascript
• node.js
• paypal
• rest-api