Repayd

Inspiration

AI agents are increasingly able to hold wallets and make payments, but a mistake, exploit, or hijacked instruction can put their owners’ funds at risk. We wanted to bring the protections people expect from financial services—transaction checks, fraud evidence, and insurance—to autonomous agents.

What it does

REPAYD protects an agent’s wallet with three transaction paths: routine payments go through, suspicious ones are held for review, and clear policy violations are blocked. If a covered transaction still gets through, REPAYD can pay a claim from its mutual pool. A cryptographic instruction history helps establish whether the owner or an outside source initiated the action, and an agent’s behavior can inform its coverage price.

How we built it

We built the system from smart contracts, a deterministic risk and verdict engine, an agent SDK, a coverage API, and dashboards. The contracts handle wallet policies, transaction holds, verdicts, and payouts. The engine evaluates transaction signals using explicit rules; it does not rely on an AI model to make coverage or payout decisions. We also built a risk subgraph to index protocol activity and support pricing and agent records, with ERC-8004 for agent identity and reputation.

What we learned

We learned that agent insurance needs more than a pool of funds. It needs clear rules for what is covered, reliable evidence about who instructed the agent, and an auditable path from a transaction to a verdict and payout. We also learned that keeping decisions deterministic makes them easier to reproduce and verify.

Challenges

A key challenge was coordinating off-chain monitoring and verdicts with on-chain wallet controls, while keeping funds recoverable during holds and preventing duplicate or excessive claims. We also had to make the same risk and pricing data usable across the contracts, API, subgraph, and dashboards. We addressed these by defining shared rules and invariants, testing the core flows, and verifying a live demo on Arc testnet.