Sentinel

Sentinel is a CRE-powered consensus layer for AI agent decisions. AI agents are moving billions onchain — trading, bridging, staking — but nothing verifies whether the agent's decision was sound before execution. One hallucination drains a wallet.

Sentinel solves this: before any onchain action executes, the Chainlink DON independently evaluates it. Multiple nodes query an LLM with the proposed action, each receives an independent risk assessment (score 0-100, risk level, specific factors), and CRE consensus produces a verified safety score. The SentinelVault smart contract then approves or blocks execution based on the consensus score. Every decision is recorded onchain with score, risk level, and reasoning hash.

The demo features an autonomous AI agent (Tidus) that scans LLM-generated market opportunities, reasons about risk/reward in real-time, and proposes trades. The DON evaluates independently — sometimes agreeing, sometimes catching risks the agent missed. All transactions are real on Sepolia.

CRE workflow (workflow/handler.ts): HTTP trigger receives trade proposals, HTTPClient sends structured LLM queries to OpenAI from within the DON, consensusIdenticalAggregation ensures all nodes agree on the risk score, runtime.report generates a signed verdict, and EVMClient writes it to SentinelVault on Sepolia.

Smart contract (SentinelVault.sol): Foundry-built vault that stores trade proposals, receives DON-attested verdicts, and executes atomic token swaps only when approved. 9 passing tests.

Server (server.ts): Express proxy with two SSE endpoints. /api/propose generates market opportunities via OpenAI and streams agent reasoning. /api/evaluate runs the full pipeline: proposeTrade onchain → CRE workflow simulation → reportVerdict onchain → executeTrade onchain (if approved).

Dashboard (React + Tailwind): Real-time visualization with staggered opportunity card reveals, streaming dialogue, DON node scores, consensus gauge, and Etherscan transaction links.

Agent roster of 10 named counterparties with DALL-E 3 portraits. Market opportunities are fully LLM-generated — nothing hardcoded.