AAC: business authorization across AI agents
A two-agent trip reservation: an $8,000 spending limit, an overreach refusal and a verified receipt. Run the demo.
Transcript
When one AI agent sends work to another, the request should carry a business authorization that states what the next agent may do—and the limits it must respect.
Agent Authority Cloud, or AAC, is a service for enforcing business authorization across collaborating AI agents. Each agent's paired sidecar verifies the authorization and its limits before the application acts.
For example, Marc works at Vantis Equity, a private equity firm. He approves a $10,000 travel budget for his trip from Austin to Shanghai. Vantis's AI planning agent works with Tourfedia's AI travel booking agent.
The AI agents handle planning and booking. Their paired AAC sidecars sign outgoing authorizations and verify incoming ones.
Vantis's planning agent limits this reservation to $8,000. It sends the request to Tourfedia's booking agent with that spending limit attached. The authorization is addressed specifically to that booking agent.
The sidecar paired with Tourfedia's booking application checks the authorization and the agent identities before passing the request to the application.
An attempted delegation tries to authorize spending $9,500, even though Tourfedia received an $8,000 limit.
The sidecar paired with Vantis rejects that delegation before the planning application runs.
The response is HTTP 403, with an error indicating that the authorization chain is invalid. An agent cannot delegate more than it was authorized to do.
The authorized $8,000 reservation succeeds. Tourfedia returns a signed receipt, and Vantis verifies that it came from the expected agent and belongs to this authorization chain.
The Agent Execution Graph brings these records together: who authorized the work, how Vantis limited it, which agent received it, and the verified receipt.
Your agents keep their business logic. Their paired AAC sidecars verify the business authorization behind each request, and leave records you can inspect afterward.
Run this two-agent demo on macOS or Linux with Python and Docker. The link is in the description.
Example reservation; supplier integration and payment omitted.