· via dev.to (home feed)
United terms reportedly allow canceling AI-agent bookings; the missing piece is scoped delegation
United's updated terms reportedly allow canceling tickets booked by unauthorized AI agents. The deeper issue: password-based agent access cannot express scope, and bans will keep coming until delegation tokens exist.

What United changed
According to a dev.to post published on November 21, United Airlines revised its Contract of Carriage on November 11 so that AI agents operating without airline authorization cannot book, change, or service tickets — even when the passenger is present and approves each action. Tickets created by such agents can be canceled without notice, the post reports. The author, who builds agent tooling at Vectle, argues this should be read as an architecture review for anyone shipping agents, not as a legal curiosity.
Passwords cannot express scope
Most agent delegation today still runs on credentials the user shares with the agent, and that, the post argues, is the real problem United is reacting to. A password is all-or-nothing: it cannot say "read only," or "book flights but never touch frequent-flyer miles," or "stop before the legal declaration at check-in." It also cannot separate the account holder, the agent acting on their behalf, and a credential-stuffing bot, because all three present the same secret. A service that accepts a password from agent traffic is effectively handling three different actors as one.
Stealth is the tempting, losing response
The predictable reaction is to make agents harder to detect — more human-like browsing so a ban cannot catch them. The post warns that United's policy already pushes in this direction: cloud agents that identify themselves get blocked, while the same delegation running inside the user's own browser goes unseen. A policy with that asymmetry discourages honesty and rewards camouflage. And a plan that depends on not being detected is not a plan at all, the author writes — it is an arms race that ends the next time the terms change.
The proposed fix: signed, scoped, revocable delegation
The workable pattern, according to the post, is one developers already apply in their own tooling: expose only the specific access a task needs, and only when it is explicitly declared. The author cites an example from Vectle's setup, where Playwright tests run through BrowserCat's cloud browser fail against localhost until a single connect option exposes loopback networking. The lesson is the shape of the fix — a single explicit setting that opens exactly the intended capability, rather than a shared credential or blanket network access.
Scaled up to travel, that becomes a delegation token: signed, scoped, and revocable, stating which agent, which passenger, and which actions are permitted — book flights under a set price, no miles, no check-in declarations — plus an expiry date. The airline verifies the grant instead of guessing who is acting. The passenger revokes it in one place. Liability becomes a field in the token rather than a dispute afterward.
United is not alone
The post situates United's move in a broader wave: Amazon blocked a shopping agent in September, US banks blocked agents in October, and Canadian banks followed. Each ban, the author writes, is a system saying the same thing — it cannot tell an authorized agent apart from an attacker, so it blocks both. Until airlines and banks accept verifiable grants of limited authority, more bans should be expected.
Why it matters
For teams building agents that transact on third-party services, contract-of-carriage language is now an architectural constraint, not fine print. A product whose core flow depends on a shared password, or on automating a human-facing interface, can be invalidated by a single terms update — and the stealth alternative turns product engineering into a detection-evasion exercise. What is missing is a standard, machine-verifiable way for users to grant agents limited authority. The builders best positioned for the coming restrictions are those designing for narrow, explicit, revocable scope from the start. The post closes by asking builders which scope they would request first, and what they would never delegate.
- #ai-agents
- #authentication
- #access-control
- #automation
- #airlines