How much freedom should an AI agent have over your software?
An agent can prepare a change, write a test or review a modification. The permissions it needs depend on the environment. In development, it can work autonomously within an isolated scope using test data. In staging, controls tighten: the change must pass tests and review before it can proceed. In our approach, a person authorises the production release.
The harness is the set of controls that enforces those rules. It defines which tools, data and actions are available. An agent changing a reminder rule can check fictional scenarios in development. Staging is where the team verifies the version and its effects. Sending a reminder to real customers requires a different level of authorisation.
During a pilot, ask to see a change that goes through and one that is blocked. Check who approved it, where the data was held and how the team can restore the previous version.
Keep reading
AI agent or conventional automation?
If inputs are structured and the rule is stable, test an explicit rule first. An agent may help with unstructured input.
How do you check code produced by several agents?
Review each change against the original need and a case the agent did not choose. A reminder may work normally but send twice after a retry.
Does software built with agents need AI in production?
An agent can write portal code without the deployed portal calling a model. An automatic summarisation feature may introduce that dependency.