Building the authority layer for AI agents.
Arx by EDGE² defines, enforces and proves what an agent may do, on whose behalf, for what purpose and for how long.
Agents act on real systems. The model cannot be the authority.
An agent does not just answer. It reads records, calls tools and alters systems of record. Once a model can act, the question is no longer only whether its output is acceptable. It is whether the action was authorised, for this task, on behalf of this person or organisation, at this moment.
A model can be trained to recognise a boundary and still be persuaded, confused or simply mistaken. Training can improve how an agent behaves. It cannot determine what the agent is authorised to do. That decision belongs outside the agent.
Principles for agent authority.
Existing identity and access controls remain the outer boundary. Arx narrows what those controls permit to the authority required for the task at hand.
Identity sets the outer limit. Arx sets the terms.
Government architecture and assurance experience, applied to AI agents.
EDGE² draws on enterprise architecture and assurance work on UK government data platforms, including applying the Government Digital Service (GDS) Service Standard and taking services through formal governance, assessment and assurance. In that environment, access to sensitive information must be justified, constrained and auditable, and services must demonstrate that they meet defined standards.
Arx applies the same discipline when AI agents act on organisational systems. Capability determines what an agent can do. Authority determines what it may do. Evidence shows what actually happened.
Tell us about a workflow where authority matters.
If you are planning or developing agents that will act on organisational data, tools or decisions, we would like to hear from you. We also welcome conversations with partners funding or researching the same problem.