How Much to Delegate
If you make the run rules stricter, can you say the agent will not do anything dangerous? You cannot. In the Marketplace edition, run rules take effect in some situations and not in others, and where the boundary lies does not match intuition.
This page explains what the run rules stop and how far, and where responsibility passes to you. The setup procedure is in the Execution Rules and Network Access Guide.
The platform does not stop operations inside a registered agent one by one
A registered agent is a container image built with the SDK. The platform does not watch what it calls or writes inside it and stop each operation. Writing something as a prohibited operation does not always stop the operation while it runs. Prohibited operations added in a project are not passed to the agent.
Run rules are fixed to the run before launch
Run rules work in two tiers. The tenant-wide rule (an administrator decides it on “Execution rules” in the admin console) and the project's rule are combined, and the stricter one becomes that run's basic rule (Open, Guarded, or Strict).
In a workflow, before a registered agent is launched, the platform checks whether that basic rule is among the “supported run rules” in the agent's certification. If it is not, the agent is not launched and the step ends as “Denied by rule”. Prohibited operations (the tenant-wide ones and the ones added in a project) are also combined before launch and fixed as the rules that apply to that run. When a prohibited operation points to a connector that does not exist, the agent is not launched. The version of the fixed rules and the basic rule are kept in the execution record.
The platform does not stop the operations of a run one by one while it runs.
If you make the project's basic rule stricter, agents that do not support that basic rule stop running. In a cycle, the basic rule has no effect.
Outbound traffic is narrowed for the whole environment
How far an agent can communicate outside is decided by settings for the whole environment, not by the project. By default, HTTP(S) to public destinations goes through the egress gateway, and destinations you want to block are registered under “Banned URL” in the admin console. Other destinations are narrowed with a NetworkPolicy.
What stops is traffic that follows these settings. For a destination you must stop for certain, block it with a NetworkPolicy as well as a banned URL.
The boundary of responsibility
What run rules and traffic limits judge is only whether it may launch and whether it may connect.
Whether the work the agent does is right for the business, and whether the result is usable, the platform cannot judge. What to delegate is decided by you, and run rules and traffic limits are a mechanism to hold to that decision mechanically. So it is not “safe because I made the run rules strict”. Prevent operations you do not want through how the agent is built and through the traffic limits.
What to read next
- Execution Rules and Network Access Guide — How to check the basic rule and narrow traffic
- Human Review — What a review step decides
- Records and Metrics — How “Denied by rule” is recorded