Morrow AI

Inspiration

AI agents can already reason, plan, and produce work, but the moment an agent needs to pay for a service, a human usually has to step in and click a payment button. That gap prevents agents from operating as accountable economic actors. Morrow AI explores a safer model: Gemini evaluates evidence, a deterministic policy layer constrains the decision, Circle Testnet executes an approved transfer, and every stage is written to an audit log.

What it does

Morrow AI is a technology demonstration of governed autonomous payments. It is not a production payment system and it does not represent real customer transactions. No production funds are used. All demonstrated transfers use Test USDC on Arc Testnet.

The workflow has four independent stages. First, Gemini reviews the task, delivery evidence, requested amount, and payment context. It returns a structured decision with confidence, rationale, and an evidence-satisfied flag. Second, the Policy Gate applies rules the model cannot override: evidence must pass, the vendor must be allowlisted, the amount must be no more than 25 Test USDC, and the network must be Arc Testnet. Third, only when every rule passes does the server call the Circle developer-controlled wallet API. Fourth, the system stores separate decision, policy, execution, and audit-seal records.

The safety behavior is as important as the successful path. Missing evidence, a Gemini error, an unapproved vendor, or a request above the autonomous limit fails closed. A blocked request never invokes Circle.

Verified result

The integration was verified with a real Arc Testnet transaction transferring 1 Test USDC between two Circle developer-controlled wallets. Circle returned HTTP 201 and transaction ID b7728847-f3ee-5c6b-94cf-6847107e3538. The final Circle state is COMPLETE. The on-chain transaction hash is:

0xae4116dadcf2444380b78fb8b195f4077f195deb68b8efbfe0280b00cbc48b89

This transaction was a one-time integration verification during setup. The application runtime implements the automated decision-to-policy-to-transfer path shown in the demo. The evidence package includes a sanitized JSON record and a Circle Console screenshot; it contains no API keys or Entity Secret.

How we built it

The server is written in TypeScript with Next.js-compatible API routes. Gemini supplies the structured payment decision. The deterministic Policy Gate remains separate from the model so spending limits and allowlists cannot be changed by prompt output. Circle's Wallets API supplies the Testnet settlement layer. Each Circle request retrieves the entity public key and encrypts the 32-byte Entity Secret with RSA-OAEP SHA-256, producing a fresh ciphertext for every transfer request. Secrets remain server-side.

Durable records distinguish actor types: AI for the Gemini decision, policy for deterministic controls, agent for payment execution, and system for the audit seal. Logs contain timestamps, model name, HTTP status, latency, provider state, transaction ID, and transaction hash when available.

Human and AI responsibilities

The human defines the allowed vendors, autonomous spending limit, network restriction, and evidence standard. Gemini assesses whether the submitted evidence supports payment. Deterministic code makes the final policy decision. The agent executes only permitted testnet transfers. Humans also review logs, rotate secrets, and decide whether the system should ever advance beyond a technology demonstration.

Challenges and lessons

The hardest part was not sending a token; it was making the action governable and provable. Circle requires a unique encrypted Entity Secret for every request. We also needed to separate probabilistic model judgment from deterministic financial rules, preserve an honest distinction between testnet and production, and create evidence that a reviewer can understand in three minutes.

What's next

The next technical steps are webhook-based confirmation, transaction-status polling, stronger recipient identity controls, policy versioning, and independent log integrity verification. Any production use would require legal, security, compliance, and operational review. For this submission, Morrow AI remains exactly what the opening card states: a technology demonstration of the future AI-native economy.

Built With

• ai-agents
• arc-testnet
• audit-logs
• circle-wallets-api
• cloudflare-workers
• d1
• gemini-api
• next.js
• test-usdc
• typescript