AI Engineer hosted WorkOS product manager Isabelle Williams as she demoed Waypoint, a travel app from a company trusted by over 3,000 customers including OpenAI and Cursor, that lets AI agents sign up via an off.md file and call APIs with scoped agent tokens instead of human credentials.
I think she nailed the core fix, because handing an agent your user token was never sustainable.
Williams showed two on-ramps, an anonymous try where the agent gets an access token plus a one-time claim code the human later redeems, and a zero-click path where the agent presents an ID-JAG, an Identity Assertion Joined Authorization Grant signed by the agent provider that carries the user's name, verified email and audience and is verified by Waypoint before it mints its own access token.
For actions that move money, she issued agent tokens whose scopes are the intersection of Jordan the approver's permissions and a blueprint ceiling, so a remote agent with a stolen token still cannot raise the budget or change approval rules without the human's user token.
That markdown-for-agents pattern is already spreading, with Expo docs noting you can append .md to any documentation URL for AI agents, which makes Williams' off.md bet feel less exotic than it sounds.
Conceded, this still asks app teams to run an authorization server that validates ID-JAGs and enforces revocation correctly, but if you are building for agents now I would ship off.md for discovery and never give an agent the user's token.
