Every action passes one gate
Before an agent emails a customer, issues a refund or changes a record, the gate checks it. The answer is allowed, needs approval or denied, with the reason.
- 01 Kill switches
- Is this agent, tool or team stopped?
- 02 Arguments
- Are the arguments allowed for this tool?
- 03 Delegation
- Does the person it acts for allow this?
- 04 Budgets
- Is there budget left for the agent, team and task?
- 05 Rules
- What do your rules say about this action?
- 06 Risk against trust
- Is the risk within the agent's trust tier?
Six checks, in a fixed order
The gate runs before anything happens, for every agent: ready-made, your own and the ones you already use.
A kill switch stops an action before anything else is considered. Risk against trust comes last, after your rules.
The answer always carries a reason. When an action is denied, the agent and the log both say why.
Example: a $240 refund clears kill switches, arguments, delegation and budgets. Then your rule that refunds need an admin makes it wait for one.
Limits that hold before the call
Budgets and kill switches are checked at the gate, so they stop the next action before it runs.
Budgets per agent
Budgets per team and task
Kill switches
Rules as code
Data classes
Frequently Asked Questions
Put an agent on real work, with people in charge
Design partners run one real workflow on OrchKernel and work directly with the engineers building it.