Sardis Whitepaper

Abstract

As AI agents evolve from conversational assistants to autonomous actors capable of executing complex workflows, they encounter a critical limitation: the inability to transact financially. Current payment infrastructure is designed to block non-human actors through mechanisms like 2FA, CAPTCHAs, and behavioral analysis.

Sardis introduces a novel financial infrastructure layer that provides AI agents with non-custodial MPC wallets, natural language spending policies, and a real-time policy engine that prevents "financial hallucinations" -- the accidental over-spending or mis-spending of funds due to agent logic errors.

The Problem

The Read-Only Trap

We are transitioning to an Agentic Economy where AI agents perform increasingly complex tasks on behalf of humans. Yet these agents remain fundamentally "read-only" when it comes to financial transactions. They can browse, plan, and recommend, but they cannot pay.

Financial Hallucination Risk

Most discourse around AI safety focuses on text hallucinations. We argue that "financial hallucination" -- an agent accidentally spending $10,000 instead of $100 due to a retry loop, decimal error, or logic bug -- represents a more immediate and quantifiable risk.

Example: The $10K Bug -- An agent tasked with purchasing $100 of API credits enters a retry loop after a timeout. Without spending limits, it executes the same purchase 100 times, resulting in $10,000 in charges.

The Solution

Financial Firewall

Sardis implements a "Financial Firewall" that sits between agents and payment rails. Every transaction is validated against a set of programmable policies before execution.

Natural Language Policies

Unlike traditional payment systems that require complex rule configurations, Sardis accepts policies in plain English:

"Allow SaaS vendors up to $100 per transaction"
"Block all retail purchases"
"Maximum $500 daily spend"
"Only allow payments to openai.com, github.com, vercel.com"

Non-Custodial Architecture

Sardis uses Multi-Party Computation (MPC) via Turnkey to ensure that agents control their own wallets without any single party having access to the complete private key.

Market Opportunity

The "Machine Customer Economy" is projected to reach $30 trillion by 2030 (Gartner). As agents become capable of autonomous purchasing decisions, the need for secure, programmable payment infrastructure becomes critical.

Technical Architecture

Sardis consists of four primary components:

1. Policy Engine -- Real-time transaction validation with natural language rule parsing. 12-check pipeline, fail-closed.
2. MPC Wallets -- Non-custodial key management via Turnkey infrastructure.
3. Settlement Layer -- Multi-rail support: on-chain (Base + CCTP v2 bridging), fiat (virtual cards, ACH), and checkout (Pay with Sardis).
4. Integration SDKs -- Native support for LangChain, Vercel AI, OpenAI Agents SDK, CrewAI, MCP (52 tools), and 10+ more frameworks.

Business Model

Sardis follows an Open Core licensing model:

Open Source:

• Python SDK (pip install sardis)
• TypeScript SDK (npm install @sardis/sdk)
• MCP Server (52 tools)
• Integration adapters (13 frameworks)

Proprietary:

• Policy Engine core
• MPC node management
• Compliance infrastructure
• Enterprise features

Conclusion

Sardis provides the missing financial infrastructure layer for the Agent Economy. By combining non-custodial wallets, natural language policies, and a real-time policy engine, we enable agents to transact safely while preventing the financial hallucinations that would otherwise make autonomous agent spending too risky to deploy.

Contact: contact@sardis.sh | GitHub