​A2Z Agentz: Autonomous Web3 Scavenger

A2Z Agentz is a fully autonomous, dual-brain AI agent system purpose-built for the AMD-powered Web3 stack, targeting the Base network. In crypto, speed and accurate intelligence are everything — so we split the workload across two specialized agents to maximize both reasoning depth and execution speed. Agent A (The Scout) is the intelligence hub. It runs Meta-Llama-3.1-8B-Instruct-AWQ-INT4 locally on AMD Instinct MI300X via vLLM + ROCm, exposed securely through a Cloudflare tunnel. Agent A performs deep OSINT signal detection — continuously scraping and analyzing DexScreener and Farcaster (Neynar) to score market opportunities in real-time using strict JSON-mode reasoning, keeping end-to-end latency under 25 seconds. Disclaimer on Technology Tags: Please note that while Agent B is powered by the newly released DeepSeek V4 Pro for superior execution logic, the platform's technology dropdown only offered "DeepSeek V3" as the closest available option. We selected the V3 tag to reflect our ecosystem choice, while actively utilizing the V4 Pro API in our live production environment. Agent B (The Vault) is the secure execution engine. Connected via a database queue, it receives high-confidence signals from Agent A and uses Fireworks llama-v3p3-70b-instruct to make split-second, security-gated on-chain execution decisions. The system ships a production Vercel dashboard with real-time WebSocket + polling fallback for live agent logs, JWT / API-key authenticated backend, and a manual Circuit Breaker to pause autonomous on-chain execution. Agent B executes real native transactions on Base mainnet with operator-configurable spend caps (hard $2.00/tx autonomous limit) and optional GoPlus security signals. Built entirely on AMD infrastructure for the AMD Lab Lab hackathon — from MI300X inference to Cloudflare tunnel edge delivery.