Developers Share Lessons From Building Fail-Closed Verification Tools for AI Agents

The author reports building two small open-source verification tools for the emerging AI agent economy, which share a rule: an unverifiable result should never be reported as clean. Al-Muqarin checks claims against signed, dated snapshots of public sanctions lists and returns a not-measured result when a list is missing, stale, or below a minimum version. agent-receipt-verify checks x402/EIP-712 payment receipts and reports incomplete ones as failures rather than passing them. The team tests its tools by planting deliberate bugs and confirming that tests catch each one, with Al-Muqarin catching all 548 planted faults. The authors acknowledge that other services already sign sanctions-screening results and that their tools have no paying users yet. They also disclose an unresolved weakness in a separate prototype, where the verifier identity field is free text and can be exploited.