CaseRelay

Problem

AI workflows that touch sensitive data create a governance gap: when should the agent act autonomously, and when must a human explicitly authorize the next step? Without a visible checkpoint, human authority becomes implicit rather than enforced.

The design question: how do you make human authority both visible and technically enforceable in an AI workflow that handles sensitive cases?

Product concept

A permission-gated AI workflow for sensitive customer cases. The system intakes a case, scopes the agent's access to only what is needed, presents a structured human approval checkpoint, and executes a protected action only after explicit authorization — with a traceable receipt.

Built for the Terminal 3 Agent Auth SDK hackathon track: Best Agent utilising Terminal 3 Agent Auth SDK.

How the system works

• Case intake: a sensitive customer case is received and categorized
• Scoped access: the agent is given access only to the information and tools required for this specific case
• Human approval checkpoint: a structured decision point is presented — no protected action can occur without explicit human sign-off
• Protected execution: only if authorized does the agent perform the protected action
• Receipt generation: a traceable record of the authorization and the action taken is produced

Main conceptual workflow

The diagram below is a conceptual illustration of the system flow. It does not reproduce internal documentation.

My contribution

• Designed the permission-gated workflow architecture and the human approval checkpoint
• Integrated with the Terminal 3 Agent Auth SDK for agent governance
• Deployed to Cloud Run for the safe-preview environment
• Submitted with frozen commit 7b85a14116f1fedf71c989fce675c618efa667ff (tag: submission-final)

Technical implementation

• Terminal 3 Agent Auth SDK — agent governance and authorization framework
• Cloud Run — containerized deployment for the safe-preview environment

Evidence and validation

• Public repository available for code inspection
• Recorded demonstration available (link below)
• Safe-preview deployment exists at the URL below — this is a demonstration environment, not a production system
• Submission frozen at commit 7b85a14116f1fedf71c989fce675c618efa667ff (tag: submission-final)

Limitations and transparency

• Safe preview: the live deployment is a demonstration environment — not a production system
• No real sensitive customer systems: the workflow does not connect to real customer data, live credentials, or production authorization infrastructure
• No production authorization: the permission gate is demonstrated, not deployed in an operational context
• Submitted / frozen: no further development is planned following submission