Put AI agents under your control.

Give an IT agent the tools it needs for a task. Keep role changes, deletions, and unrelated services outside its authority.

Let routine work run automatically and require your operator’s approval for production. A prompt can’t expand the permissions you set.

Choose what your agent can do

policy "agent-actions" {
    allow when {
        request.operation == "restart"
        request.service == "billing-api"
        request.environment in ["staging", "production"]
    }

    require operator if (request.environment == "production")
}

Review the recipient and message before you authorize sending.

Let agents draft and send customer emails while you keep approval over the recipient and message. If either changes after review, require a fresh approval.

Keep control of what gets sent

authority "outbound-message" {
    intent = "custom"
    version = "1"

    input request {
        operation: Text
        recipient: Text
        subject: Text
        body: Text
    }

    approver operator {
        type = "public-key"
        id = "operator"
        algorithm = "ED25519"
        publicKey64 = "T9U4X0BiYmULPIhbIU8/Xs/qNOhY516SF35rDUevc+w="
    }

    policy "outbound-message" {
        allow when {
            request.operation == "email.send"
        }

        require operator
    }
}

Delegate one action at a time and choose how long that permission stays valid.

Keep control of the action, target, and parameters you authorize. Whoever holds the mandate can use it only for the original operation, within its lifetime.

Delegate one specific action

1. Set the task

   Get a mandate for one exact action, including the target and parameters.

2. Keep changes out of scope

   A different action, target, or parameter requires a new mandate.

3. Choose its expiry

   Set how long the permission stays valid and collect the required approvals before use.