OAuth for AI agents: does it fit?

OAuth is excellent for "let this app act for this user." It's a poor fit for "let this agent spend up to $20 reading these repos for an hour, and let it sub-delegate."

What OAuth gives you

OAuth 2.0 issues access tokens after a consent flow, scoped by coarse scopes and validated (often) by a call to the provider. It's the right tool for delegated user access to an API. Nothing about it is wrong — it just wasn't designed for autonomous, money-spending, self-delegating programs.

What agents additionally need

• Fine, per-task scope — resource globs and actions, not a fixed scope list.
• Stateful limits — cumulative spend and rate caps enforced across calls.
• Offline attenuation — an agent narrows and sub-delegates without calling the issuer.
• Holder-binding — a stolen token is useless (proof-of-possession).
• Provable audit — a tamper-evident log of every decision.

Warrants are built around exactly these. You can use OAuth to authenticate the human/app that issues warrants, then hand agents warrants for the actual work.