Hi HN, I’m a software engineer who’s been building AI agents for about 1.5 years.I’ve noticed that teams often give agents the ability to perform real tasks, such as refunding money, changing accounts, and writing to production databases, without involving a human in the process. When there is an approval step, there’s no reliable record of who approved what or when.To address this issue, I created Approv. It’s a simple API call that pauses an agent’s risky action and prompts a human to approve or reject it via WhatsApp (with an SMS fallback). Every state change is hashed and signed with Ed25519, ensuring that the audit trail is tamper-evident and can be independently verified by anyone. The response includes the public key and the exact hashing formula.The stack consists of Deno edge functions (with a hexagonal core), Postgres (using pgmq queue and pg_cron), Twilio, and a Next.js dashboard. The app is live and free to try.Here are some links for further information:- Live app: https://approv-app.vercel.app - 90-second demo: https://www.loom.com/share/1b286fb19b1a4730a2d5ed6e8a9c2ac3I would genuinely appreciate your feedback on the audit design and whether the human-in-the-loop model aligns with your approach to building agents. Feel free to critique it thoroughly.