Governed Agent

Agents differ from assistants in consequence: an incorrect answer becomes an incorrect action affecting systems and people outside the conversation. Governing them requires the same apparatus as governing a person — a statement of what they may do, boundaries on what they may reach, confirmation before irreversible steps, and a record of what they did. The common failure is granting capability and calling it authority, which leaves the organization with automated actions nobody approved.

Audiences and roles bound what any process in Centralpoint can reach, Data Triggers make the conditions for automated action explicit rather than implicit in code, and actions are recorded alongside the AI activity that prompted them. An agent's behaviour is reviewable against the authority it was granted rather than against an impression of whether it seemed reasonable.