# Revocation

> Pull an agent's authority mid-flight — a credential, a grant, a running session — and have it take effect on the very next action, not the next restart.

Source: https://openap.org/docs/revocation

Sometimes the answer is "stop — now." A credential leaked, a grant was a mistake, a session is doing something
it shouldn't. OAP's authority is designed to be **pulled in flight**: revoke it and the change
lands on the agent's next action, not whenever it happens to restart.

## Revocation is checked at the choke point, live

Because every consequential thing an agent does is a tool call through a [gated pipeline](https://openap.org/docs/agent-loop), there's
a single place to enforce a withdrawal. Revocation re-checks **immediately before the call runs** — so a
credential or grant you pull mid-turn stops the *very next* tool call, rather than being caught only when a
token is next minted or a pod next starts. There's no window where authority you've already removed keeps
working.

## What you can pull

- **A credential** — a leaked or rotated secret. Once revoked, tools that authenticate with it fail closed, and
  repeated auth failures [end the session](https://openap.org/docs/budgets-breakers) rather than retrying a dead credential.
- **A grant** — an approval or a slot you granted, including a [participant's standing](https://openap.org/docs/multiplayer-sessions)
  in a session. Revoke it and that subject can no longer drive or approve.
- **A session** — halt a run outright, or place it on a **hold** so it's frozen for inspection rather than
  torn down.

> **Security: Removed means removed**
>
> The value of an authorization model is only as good as how fast you can take authority back. OAP's answer is
> "at the next check," and the next check is never far away — it's the same per-action gate that authorized the
> agent in the first place. Nothing keeps running on authority you've already revoked.
