Creator case study · Haisam Arshad
Putting a human decision between a tool call and its effect
AgentGuard uses a refund workflow to answer a concrete question: if an AI agent requests a financial action, who can actually authorize it, and how can an operator verify what happened?
Identity is only the first check
Auth0 identifies each runtime. The MCP server checks permissions before evaluating the request. A shared budget stops smaller requests from bypassing the spending limit, and higher-risk requests wait for the operator.
Retries must not create extra effects
An agent can lose a response and retry. The request ID and stored payload let the database recognize the same request. Review and its audit event commit together; execution, receipt creation, and its audit event also commit together.
Keep the evidence honest
The deployed test recorded a human-approved $755 simulated refund once, then rejected replay. The reference application keeps the payment effect in a sandbox so reviewers can inspect outcomes without moving money. The report identifies the precise builds, the trust boundary, and the tests still to run.