AgentEscrow

Inspiration

I wanted to solve the massive amount of friction that exists in B2B micro-contracting and freelance work. Freelancers and small businesses waste endless hours haggling over scopes, chasing milestone approvals, and dealing with payment delays. Inspired by the rise of autonomous agent workflows, I set out to ask: What happens when AI agents take over the negotiation, contract execution, and financial settlement process entirely, using PayPal as the trusted rail?

What it does

AgentEscrow is an autonomous AI agent system where two business agents (a Buyer Agent and a Vendor Agent) negotiate a project scope, break it down into micro-milestones, lock funds securely into PayPal sandbox APIs, and automatically release payouts as the AI independently verifies deliverables. Users get a split-screen dashboard where they can watch the live agent negotiation log and track the escrow ledger in real-time.

How I built it

I built AgentEscrow using a modern tech stack focused on speed and reliability:

• Frontend & Backend API: Built using Next.js (App Router) with TypeScript and Tailwind CSS for a sleek, dark-mode "AI Command Center" aesthetic.
• Agent Logic: Integrated LLM handler routines to power the multi-agent negotiation loop between the Buyer and Vendor personas, alongside an AI Evaluator to check code or text deliverables.
• Payment Integration: Integrated official PayPal REST API endpoints (https://api-m.sandbox.paypal.com) for handling sandbox orders and automated payouts.
• Resilience: Built a robust mock-service toggle so the application functions seamlessly even if sandbox network latency spikes during testing or demonstration.

Challenges I ran into

Managing asynchronous states between multi-agent LLM chat loops and financial transaction triggers was tricky. Ensuring that the UI cleanly transitioned states from Negotiating to Escrow Locked to Verified & Paid Out without breaking required careful state management. Additionally, handling PayPal sandbox API authentication tokens while keeping the app responsive demanded a clean fallback mock mode.

Accomplishments that I'm proud of

• Successfully creating a fully functional multi-agent loop that translates a natural language prompt into a structured JSON contract.
• Integrating real PayPal REST API calls for automated milestone payouts.
• Designing a gorgeous, futuristic dashboard that clearly visualizes autonomous financial flows in real-time.

What I learned

Deepening my understanding of PayPal's developer ecosystem—specifically how Orders and Payout APIs interact with programmatic backend triggers—was incredible. I also learned a lot about structuring multi-agent system prompts to negotiate effectively and output reliable structured data schemas.

What's next for AgentEscrow

• Expanding the agent capabilities to support multi-party dispute resolutions.
• Adding support for GitHub and Figma webhooks so deliverables are verified automatically based on code commits or design file updates.
• Deploying a fully polished public version for independent creators worldwide.

Built With

• agents
• ai
• api
• css
• llm
• next.js
• node.js
• paypal
• react
• render
• rest
• sandbox
• tailwind
• typescript