AgentOG is a human approval and execution-control layer for high-impact AI agent actions.
It sits between AI agents and real-world systems like payments, bookings, forms, APIs, and external workflows. Before an agent executes an action, AgentOG creates an action fingerprint, gets human approval, issues a short-lived action-bound token, and blocks execution if the final payload changes.
Production demo: https://agent-og.vercel.app
Normal MFA verifies the user.
AgentOG verifies the exact AI agent action.
Instead of asking, “Do you approve this agent?”, AgentOG asks:
Do you approve this exact action, with this vendor, amount, data, and conditions?
If the agent changes the price, vendor, timing, sensitive data, or execution payload after approval, AgentOG blocks it.
User / Web / AI Agent
↓
Agent proposes action
↓
AgentOG creates action fingerprint
↓
Human approves exact action
↓
Short-lived token issued
↓
Execution gate verifies final payload
↙ ↘
Allowed Blocked
↓ ↓
Payment / Tampered payload rejected
Booking /
Forms
↓
Audit receipt
For the demo, AgentOG uses a ride-booking scenario to show the general framework.
User calls AgentPhone
↓
AgentOG receives voice webhook
↓
Request is parsed into structured action
↓
Browser Use / seeded flow selects an option
↓
AgentOG creates action fingerprint
↓
AgentMail sends approval link
↓
AgentPhone calls guardian
↓
Human approves exact action
↓
AgentOG issues short-lived token (fingerprint gate clears on approve)
↓
Valid execution allowed
↓
Tampered execution blocked
↓
Stripe sandbox opens only after approval
↓
AgentMail sends audit receipt
The ride-booking flow is only the example. AgentOG is built for any high-impact AI action.
- Agent commerce
- Payments and procurement
- Booking and scheduling
- Form submission
- Healthcare and caregiving workflows
- Accessibility support
- Refund negotiation
- External email actions
- API execution
- Sensitive data sharing
Used for voice request intake and guardian approval calls.
Used for browser automation and option selection.
Used for approval links and audit receipts.
Used for simulated payment execution after AgentOG approval.
Used for parsing natural language requests into structured actions and classifying high-impact actions.
Used for real-time policy and context retrieval.
Used for long-term user preferences, approval rules, and guardian context.
Payment tools help agents pay.
Browser tools help agents act.
Email and voice tools help agents communicate.
AgentOG decides whether the agent is allowed to execute.
The core value is action-bound approval: human approval is tied to the exact payload, and execution is blocked if the payload changes.
This project demonstrates:
- Voice-based agent request
- Structured action creation
- Action fingerprinting
- Human approval flow
- Short-lived execution token
- Execution verification
- Tampered payload blocking
- Stripe sandbox execution
- Audit receipt generation
AgentOG is transaction signing for AI agents: before an AI agent spends money, books a service, submits a form, or shares sensitive data, AgentOG verifies human approval and blocks execution if the final action changes.