The Ethics of Agency: Defining Rules of Engagement

How do we ensure agents remain “helpful and harmless” when they have the keys to our production infrastructure? Not with principles. With rules. And not with rules in a document — with rules in code, enforced at runtime.

The philosophical debate about AI ethics is important. But for the enterprise deploying agents today, it translates to a very concrete question: What, exactly, is this agent allowed to do?

The Social Contract for Machines

Human societies have evolved over millennia to develop implicit and explicit contracts about acceptable behavior. These contracts — laws, professional codes, social norms — exist because humans living and working together need boundaries that everyone can rely on.

Autonomous agents operating in human systems need the same thing. When an agent has access to customer data, financial systems, and communication channels, its “contract” with the humans it serves needs to be explicit, enforceable, and auditable.

We call this the Digital Social Contract — the set of rules that define the agent’s permitted scope of action, the conditions under which it can escalate to humans, and the categories of action that are categorically prohibited.

Why Abstract Ethics Fail

“Be helpful, harmless, and honest” is a beautiful principle. It is a terrible policy.

Consider a concrete scenario: an agent is given access to customer records to handle support inquiries. A customer asks the agent to look up another customer’s order history to help them “compare prices.” Is this helpful? Yes. Is it harmless? Probably not — it violates the other customer’s privacy. Is the agent honest? It’s irrelevant — the harm has already occurred.

Abstract ethical principles cannot adjudicate this scenario in real time, at the speed of an agent processing thousands of requests per hour. What is needed is a specific, enforceable rule:

rules:
  - tool: "crm.read_customer"
    action: allow
    conditions:
      - requesting_user_id: "must_match_record_owner"
    deny_reason: "Cross-customer data access is prohibited"

This rule is unambiguous. It applies consistently. It produces the same outcome every time. That is governance.

The Granularity Requirement

Effective governance for AI agents requires granularity at the level of individual actions and their parameters. It is not sufficient to say “this agent can access the database.” You need to specify:

• Which tables it can read
• Whether it can write, and to which tables
• What volume of data it can retrieve per request
• Whether it can delete records, and under what conditions
• Who can approve exceptions

This granularity is what separates a permission model from a governance model. Permission models control access. Governance models control how that access is used.

Governance as a Continuous Practice

The rules of engagement for AI agents are not a one-time specification. As agents become more capable, as business needs evolve, and as edge cases emerge in production, the policy must be updated. This is governance as a practice — an ongoing process of refinement.

The organizations that handle this well will treat agent governance the same way they treat legal and compliance work: as a permanent function with dedicated resources, not a one-time setup task.

GovernorAI by SentinelLayer provides the interface for humans to define and refine the boundaries of Agency — a policy management system that translates ethical intent into enforceable code, and a feedback loop that surfaces edge cases for human review. As AI grows more capable, it remains strictly bounded by the rules humans set.

The question is not whether to govern your agents. The question is whether to do it explicitly or to let it happen implicitly — and implicitly always means inadequately.