skip to content
Mohamed Azahrioui
← work

AI agent authorization · 2026

TrustGate

A runtime authorization gate for AI agents that checks live business evidence before a risky action runs, returning ALLOW, APPROVAL_REQUIRED, or BLOCK with an auditable receipt.

Solo · Google Cloud Rapid Agent Hackathon

snapshot
type
AI agent authorization
period
2026
source
public github
demo
watch

view repository →

problem

AI agents can propose high-impact actions such as refunds, but production systems need a transparent control layer that decides whether the action is allowed, needs approval, or should be blocked.

outcomes

  • Example verdict: a refund proposed on data synced 42 minutes ago against a 15-minute SLA, for a customer tier the input contract never authorized, returns APPROVAL_REQUIRED with the evidence receipt attached
  • The agent was authorized; the data was not trusted, and the receipt shows why
  • Connected live data evidence to policy decisions on Google Cloud

what i built

  • Runtime gate between AI agents and risky actions
  • ALLOW, APPROVAL_REQUIRED, and BLOCK decisions
  • Deterministic policy engine for final authorization
  • Auditable decision receipts instead of black-box scores
  • Live evidence ingestion from Fivetran and BigQuery
  • Data freshness checks against SLAs
  • Schema compliance checks
  • MCP endpoint deployed on Google Cloud Run
  • Small React control-room dashboard
  • Vertex AI/Gemini integration for agent proposals

tech stack

Node.jsGoogle Cloud RunVertex AIGeminiBigQueryFivetranMCPReact

Read the source

The repo has the code, structure, and implementation detail.