Decide whether a real-world action may happen now.
Before an AI agent pays, refunds, changes a record or moves a robot, Guardian checks that it does only what was approved, never causes the same effect twice, and stops when it is unclear whether an earlier attempt already happened.
Answered together, before the action takes effect.
Satisfying one never substitutes for another. A valid execution does not make an unknown outcome safe.
Within what was approved?
The request that actually goes out must match what was approved for this task. Approve a €100 refund for order 48213, and €1,000, or another order, is refused.
Is the execution still valid?
Retries, replanning and reassignment create new executions. A replaced or cancelled one cannot act, even if its credentials still work.
Has it already happened, or is it unknown?
If the same action already took effect, it must not happen again. If nobody knows yet, it waits until the facts are established.
Allow. Hold. Block.
It may go ahead.
Within what was approved, from a valid execution, and not already done or in doubt.
Find out first.
An earlier attempt's outcome is unknown. The action waits until evidence establishes what happened.
It must not go ahead.
It exceeds what was approved, its execution is no longer valid, or the same effect has already happened.
What makes it different from a retry policy.
New approvals don't skip old uncertainty.
A new task, a new execution, another agent or a fresh approval can all be valid and still wait, because an earlier attempt at the same action is unresolved.
Late answers update facts, not authority.
A confirmation that arrives after an execution was replaced can establish what happened. It never makes the old execution valid again.
"It didn't happen" needs two proofs.
Seeing no result now is not enough. It must also be established that the earlier request can no longer take effect later.
Not "each kind of action only once".
The same effect never happens twice. A new refund the customer asks for later is a different effect and goes through normally.
Use it when the action takes effect in a system you control.
Your own execution point
Robot task controllers, your own payment or banking ledger, transaction engines and enterprise automation platforms, wherever Guardian can be asked before the action takes effect.
Not functional safety
Guardian does not replace motion control, collision avoidance, emergency stops or other functional-safety mechanisms, and does not interrupt a physical action already in progress. In robotics it works at the level of discrete tasks and skills.
If the action takes effect in someone else's system, see Execution Gateway.
Choose one action to control.
Bring one workflow where an agent or planner can cause a real effect, and the failure cases that worry you. We map where Guardian should sit.