Governing the actions of an agent

Governance is based on several complementary levels:

1. the active connector makes the action technically accessible;
2. competence allows the agent to mobilise it;
3. the authorized workflow limits usable revenues;
4. permission decides whether the action is automatic, confirmed or blocked;
5. confirmation gives the last human agreement when required.

For a sensitive action, remove it from the skills if the agent never needs it. If it is sometimes useful, keep it but choose Confirmation required. Use Blocked when no route should run it.

Before publication, test the readings, then a confirmed action with fictitious or impact-free data.

Example control chain

To allow an agent to create an appointment:

1. enable Google Calendar;
2. Associate the availability reading skill;
3. Allow the appointment workflow;
4. Set event creation to Confirmation required ;
5. Publish the agent after a reading test.

Each step answers a different question: Is the tool connected, does the agent know how to use it, is the course allowed, is writing controlled and is the agent active? A single-level negative response must prevent or limit action.