All solutions

AI companies

$5,000 of model spend in an hour

Example policy threshold, not a measured loss.

Your agent holds production credentials. You hold the line.

An agent holding production credentials can spend, leak or delete at machine speed.

Live signal — agents runningThe ledger — settled

BLOCKED AT THE GATE

Agent-driven model/API spend passes the hourly cap — spend breaker
Agents spend at machine speed. Runledge is where the spend gets checked.

A limit for the consequential change.

ActionGateDecision / record
01

Credit or refund up to $500

Allowed

Allow

support agents at an AI company issue small credits constantly — friction here just slows the product's own support.

02

Credit or refund over $500

Held

Hold for the support lead

above a line, machine-speed goodwill becomes real money leaving without a human decision.

03

Agent-driven model/API spend passes the hourly cap

Blocked

Block, spend breaker [Team+]

a runaway loop burns five figures of API spend overnight; the breaker is the only thing awake at 3am.

04

A secret or API key appears in an outbound action

Blocked

Block (detection)

an agent that pastes a key into a ticket or external tool leaks the keys to the whole product.

05

Production data deletion

Held

Hold for 2 approvers [Team+]

an AI product holds its users' data; a delete is one call with no undo.

06

Instructions hidden in a customer message or ticket

Held

Hold for the support lead (detection)

every customer input is a potential prompt injection aimed at an agent holding production credentials.

07

More than 25 external actions in an hour from one agent

Held

Hold, velocity breaker [Team+]

loops are the default failure mode when agents serve agents; the rate is the tell.

08

Outbound action on an account or entity that is not the verified owner's entity

Held

Hold for support lead

an amount limit does not verify whose account is being changed; the support lead checks the target entity.

09

Outbound action carries a customer email field

Adjusted

Adjust: redact the email field

the action proceeds; the customer's email stays out of the third-party tool.

What's your scariest agent action?
Test it in shadow mode first.

Start free with the Developer plan — 10,000 actions every month. Start in shadow mode, then block actions when you trust the rules (or hold for review on Pro and above). Shadow is free during the 14-day shadow pilot for each workspace. After the 14-day shadow pilot, shadow Held, Blocked, or Adjusted outcomes count toward your plan like enforcing mode does.

Start free

Where this goes: the same rules and signed records, between companies.

Need a security review, in-VPC deployment or a custom contract? Leave your work email and we will reply within one business day.