AgentShield launches a spending firewall for autonomous AI agents, using three parallel risk checks to approve, hold, or block payments.

AgentShield evaluates agent spend requests through Financial Triangulation: three independent checks (quantitative budget/pattern analysis via Redis, policy enforcement via Postgres, and semantic goal alignment via Claude Haiku) run in parallel, producing verdicts of SAFE (immediate execution), SUSPICIOUS (human review hold), or MALICIOUS (blocked).

The system supports stablecoin spending (USDC/USDT) on ethereum, base, solana, polygon, and arbitrum networks, with optional fiat adapter compatibility. Agents authenticate via HMAC-SHA256 signed requests; humans approve or deny pending spends through an email link or React dashboard.

Built after a buying agent attempted a bad purchase that AgentShield caught, the platform uses SendGrid for human-in-the-loop (HITL) notifications and stores audit logs and transaction activity in PostgreSQL, with local development available via Docker and Python 3.11+.