Software needs a better answer than “the request returned success.”
Modern automated systems can make real changes faster than humans can review each step. The engineering problem is not only whether an action was permitted. It is also whether the intended effect actually happened, whether a retry is safe, and whether the evidence is strong enough to support the next decision.
Three questions, not one
What was intended?
Keep the action, constraints and protected intent explicit enough that later evidence can be compared against them.
What was authorized?
Bind permission to the exact operation, resource, policy, approval and one-use execution context rather than an open-ended “agent may act” assumption.
What happened?
Observe the external effect independently when possible. If the evidence cannot resolve it, preserve the uncertainty and avoid blind retry.
Why this matters now
AI agents and automated workflows make timeouts, tool retries, partial execution, stale context, delegated authority and high-speed change more consequential. Strong model reasoning does not eliminate those distributed-systems failure modes. RuleOak's engineering approach is to make the boundary explicit rather than asking the model to reason its way around an unobservable effect.
A broader engineering principle
The same discipline applies beyond agent actions: release claims should be bound to the exact evidence that supports them; user-owned state should remain understandable; recovery should be designed; and systems should fail narrower when decision-critical authority or evidence is missing.